nips nips2010 nips2010-43 knowledge-graph by maker-knowledge-mining

43 nips-2010-Bootstrapping Apprenticeship Learning

Source: pdf

Author: Abdeslam Boularias, Brahim Chaib-draa

Abstract: We consider the problem of apprenticeship learning where the examples, demonstrated by an expert, cover only a small part of a large state space. Inverse Reinforcement Learning (IRL) provides an efﬁcient tool for generalizing the demonstration, based on the assumption that the expert is maximizing a utility function that is a linear combination of state-action features. Most IRL algorithms use a simple Monte Carlo estimation to approximate the expected feature counts under the expert’s policy. In this paper, we show that the quality of the learned policies is highly sensitive to the error in estimating the feature counts. To reduce this error, we introduce a novel approach for bootstrapping the demonstration by assuming that: (i), the expert is (near-)optimal, and (ii), the dynamics of the system is known. Empirical results on gridworlds and car racing problems show that our approach is able to learn good policies from a small number of demonstrations. 1

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 ca Abstract We consider the problem of apprenticeship learning where the examples, demonstrated by an expert, cover only a small part of a large state space. [sent-7, score-0.27]

2 Inverse Reinforcement Learning (IRL) provides an efﬁcient tool for generalizing the demonstration, based on the assumption that the expert is maximizing a utility function that is a linear combination of state-action features. [sent-8, score-0.276]

3 In this paper, we show that the quality of the learned policies is highly sensitive to the error in estimating the feature counts. [sent-10, score-0.13]

4 To reduce this error, we introduce a novel approach for bootstrapping the demonstration by assuming that: (i), the expert is (near-)optimal, and (ii), the dynamics of the system is known. [sent-11, score-0.539]

5 Empirical results on gridworlds and car racing problems show that our approach is able to learn good policies from a small number of demonstrations. [sent-12, score-0.14]

6 The expert’s help consists in simply specifying a reward function. [sent-16, score-0.261]

7 However, in many practical problems, even specifying a reward function is not easy. [sent-17, score-0.261]

8 In fact, it is often easier to demonstrate examples of a desired behavior than to deﬁne a reward function (Ng & Russell, 2000). [sent-18, score-0.261]

9 apprenticeship learning, is a technique that has been widely used in robotics. [sent-22, score-0.208]

10 An efﬁcient approach to apprenticeship learning, known as Inverse Reinforcement Learning (IRL) (Ng & Russell, 2000; Abbeel & Ng, 2004), consists in recovering a reward function under which the policy demonstrated by an expert is near-optimal, rather than directly mimicking the expert’s actions. [sent-23, score-1.082]

11 The learned reward is then used for ﬁnding an optimal policy. [sent-24, score-0.308]

12 Unfortunately, as already pointed by Abbeel & Ng (2004), recovering a reward function is an ill-posed problem. [sent-26, score-0.281]

13 In fact, the expert’s policy can be optimal under an inﬁnite number of reward functions. [sent-27, score-0.579]

14 Most of the work on apprenticeship learning via IRL focused on solving this particular problem by using different types of regularization and loss cost functions (Ratliff et al. [sent-28, score-0.248]

15 IRL-based algorithms rely on the assumption that the reward function is a linear combination of state-action features. [sent-32, score-0.261]

16 Therefore, the value function of any policy is a linear combination of the expected discounted frequency (count) of encountering each state-action feature. [sent-33, score-0.326]

17 In particular, the value function of the expert’s policy is approximated by a linear combination of the empirical averages of the features, estimated from the demonstration (the trajectories). [sent-34, score-0.44]

18 For the tasks related to systems with a stochastic dynamics and a limited number of available examples, we propose an alternative method for approximating the expected frequencies of the features under the expert’s policy. [sent-36, score-0.141]

19 Our approach takes advantage of the fact that the expert’s partially demonstrated policy is near-optimal, and generalizes the expert’s policy beyond the states that appeared in the demonstration. [sent-37, score-0.652]

20 We denote by MDP\R a Markov Decision Process without a reward function, i. [sent-42, score-0.261]

21 We assume that the reward function R is given by a linear combination of k k feature vectors fi with weights wi : ∀s ∈ S, ∀a ∈ A : R(s, a) = i=0 wi fi (s, a). [sent-45, score-0.418]

22 A deterministic policy π is a function that returns an action π(s) for each state s. [sent-46, score-0.381]

23 A stochastic policy π is a probability distribution on the action to be executed in each state, deﬁned as π(s, a) = P r(at = a|st = s). [sent-47, score-0.316]

24 The value V (π) of a policy π is the expected sum of rewards that will be received if policy π will ∞ be followed, i. [sent-48, score-0.614]

25 An optimal policy π is one satisfying π = arg maxπ V (π). [sent-51, score-0.354]

26 The occupancy µπ of a policy π is the discounted state-action visit distri∞ bution, deﬁned as: µπ (s, a) = E[ t=0 γ t δst ,s δat ,a |α, π, T ] where δ is the Kronecker delta. [sent-52, score-0.353]

27 The frequency of a feature fi for a policy π is given by vi,π = F (i, . [sent-56, score-0.417]

28 Using this deﬁnition, the value of a policy π can be written as a linear function of the frequencies: V (π) = wT F µπ = wT vπ , where vπ is the vector of vi,π . [sent-58, score-0.295]

29 Therefore, the value of a policy is completely determined by the frequencies (or counts) of the features fi . [sent-59, score-0.502]

30 1 Apprenticeship Learning Overview The aim of apprenticeship learning is to ﬁnd a policy π that is at least as good as a policy π E demonstrated by an expert, i. [sent-61, score-0.82]

31 The value functions of π and π E cannot be directly compared, unless a reward function is provided. [sent-64, score-0.261]

32 To solve this problem, Ng & Russell (2000) proposed to ﬁrst learn a reward function, assuming that the expert is optimal, and then use it to recover the expert’s complete policy. [sent-65, score-0.537]

33 However, the problem of learning a reward function given an optimal policy is ill-posed (Abbeel & Ng, 2004). [sent-66, score-0.579]

34 In fact, a large class of reward functions, including all constant functions for instance, may lead to the same optimal policy. [sent-67, score-0.284]

35 To overcome this problem, Abbeel & Ng (2004) did not consider recovering a reward function, instead, their algorithm returns a policy π with a bounded loss in the value function, i. [sent-68, score-0.597]

36 V (π) − V (π E ) , where the value is calculated by using the worst-case reward function. [sent-70, score-0.294]

37 This property is derived from the fact that when the frequencies of the features under two policies match, the cumulative rewards of the two policies match as well, assuming that the reward is a linear function of these features. [sent-71, score-0.588]

38 In the next two subsections, we brieﬂy describe two algorithms for apprenticeship learning via IRL. [sent-72, score-0.208]

39 , 2006), is a robust algorithm based on learning a reward function under which the expert’s demonstrated actions are optimal. [sent-74, score-0.334]

40 (2008), is a fast algorithm that directly returns a policy with a bounded loss in the value. [sent-76, score-0.316]

41 A policy maximizing the cost-augmented reward vector (wT F + l) is almost completely different from π E , since an additional reward l(s, a) is given for the actions that are different from those of the expert. [sent-80, score-0.868]

42 This algorithm minimizes the difference between the value divergence wT F µπE − wT F µ and the policy divergence lµ. [sent-81, score-0.327]

43 (2006) showed that cq can be q minimized by using a subgradient method. [sent-84, score-0.186]

44 For a given reward w, a subgradient gw is given by: q−1 q gw = q (wT F + l)µ+ − wT F µπE F ∆w µπE + λw (3) where µ+ = arg maxµ∈G (wT F + l)µ, and ∆w µπE = µ+ − µπE . [sent-85, score-0.365]

45 3 Linear Programming Apprenticeship Learning Linear Programming Apprenticeship Learning (LPAL) is based on the following observation: if the reward weights are positive and sum to 1, then V (π) V (π E ) + mini [vi,π − vi,πE ], for any policy π. [sent-87, score-0.606]

46 LPAL consists in ﬁnding a policy that maximizes the margin mini [vi,π − vi,πE ]. [sent-88, score-0.393]

47 The learned policy π is given by: µπ (s, a) π(s, a) = a ∈A µπ (s, a ) 3. [sent-93, score-0.319]

48 4 Approximating feature frequencies def Notice that both MMP and LPAL require the knowledge of the frequencies vi,ππE = F (i, . [sent-94, score-0.291]

49 These frequencies can be analytically calculated (using Bellman-ﬂow constraints) only if π E is completely speciﬁed. [sent-96, score-0.156]

50 , sm , am , ), 1 1 H H the frequencies vi,πE are estimated as: vi,πE = ˆ 1 M M H γ t fi (sm , am ) t t (5) m=1 t=1 There are nevertheless many problems related to this approximation. [sent-100, score-0.233]

51 First, the estimated frequencies vi,πE can be very different from the true ones when the demonstration trajectories are scarce. [sent-101, score-0.344]

52 Secˆ ond, the frequencies vi,πE are estimated for a ﬁnite horizon H, whereas the frequencies vi,π used in ˆ the objective function (Equations (2) and (4)), are calculated for an inﬁnite horizon (Equation (1)). [sent-102, score-0.299]

53 Finally, the frequencies vi,πE are a function of both a policy and the transition probabilities, the empirical estimation of vi,πE does not take advantage of the known transition probabilities. [sent-104, score-0.46]

54 Consequently, the reward loss ∆w 2 approaches zero as the error of the estimated feature frequencies ∆vπ 2 approaches zero. [sent-118, score-0.429]

55 ˆ ˆ Figure (1) illustrates the divergence from the optimal reward weight wE when approximate frequencies are used. [sent-121, score-0.423]

56 However, this cannot be done unless the complete expert’s policy π E is provided. [sent-125, score-0.295]

57 In fact, it corresponds to a margin between two convex functions: the value of the bootstrapped expert’s policy maxµ∈GπE wT F µ and the value of the best alternative policy maxµ∈G (wT F + l)µ. [sent-129, score-0.726]

58 In practice, as we will show in the experimental analysis, the solution returned by the bootstrapped MMP outperforms the solution of MMP where the expert’s frequency is calculated without taking into account the known transition probabilities. [sent-131, score-0.201]

59 The computational cost of minimizing this modiﬁed cost function is twice the one of MMP, since two optimal policies are found at each iteration. [sent-133, score-0.184]

60 Proof If q = 1 and λ = 0, then the cost cq (w) corresponds to a distance between the convex and piecewise linear functions maxµ∈G (wT F + l)µ and maxµ∈GπE wT F µ. [sent-137, score-0.204]

61 Therefore, for any vector µ ∈ GπE , the function cq is monotone in the interval of w where µ is optimal, i. [sent-138, score-0.164]

62 Consequently, the number of local minima of the function cq is at most equal to the number of optimal vectors µ in GπE , which is upper bounded by the number of deterministic policies deﬁned on S\Se , i. [sent-141, score-0.287]

63 Consequently, the number of different local minima of the function cq decreases as the number of states covered by the demonstration increases. [sent-144, score-0.329]

64 Ultimately, the function cq becomes convex when the demonstration covers all the possible states. [sent-145, score-0.289]

65 Theorem 1 If there exists a reward weight vector w∗ ∈ Rk , such that the expert’s policy π E is the only optimal policy with w∗ , i. [sent-146, score-0.874]

66 arg maxµ∈G w∗ T F µ = {µπE }, then there exists α > 0 such that: (i), the expert’s policy π E is the only optimal policy with αw∗ , and (ii), cq (αw∗ ) is a local minimum of the function cq , deﬁned in Equation (6). [sent-148, score-0.977]

67 Proof The set of subgradients of function cq at a point w ∈ Rk , denoted by w cq (w), corresponds to vectors F µ − F µ , with µ ∈ arg maxµ∈G (wT F + l)µ and µ ∈ arg maxµ∈GπE wT F µ. [sent-149, score-0.4]

68 In order that cq (w) will be a local minimum, it sufﬁces to ensure that 0 ∈ w cq (w), i. [sent-150, score-0.328]

69 Let w∗ ∈ Rk 5 be a reward weight vector such that π E is the only optimal policy, and let = w∗ T F µπE − w∗ T F µ where µ ∈ arg maxµ∈G−{µπE } w∗ T F µ. [sent-153, score-0.32]

70 6 Bootstrapping Linear Programming Apprenticeship Learning As with MMP, the feature frequencies in LPAL can be analytically calculated only when a complete policy π E of the expert is provided. [sent-160, score-0.752]

71 ,k−1 minπ ∈ΠE [vi,π −vi,π ], where ΠE denotes the set of all the policies that select the same actions as the expert in all the states that occurred in the demonstration, assuming π E is deterministic (In LPAL, π E is not necessarily an optimal policy). [sent-164, score-0.507]

72 ,k−1 [vi,π − vi ], where def E vi = maxπ ∈ΠE vi,π for each feature i. [sent-168, score-0.309]

73 , k − 1} : v vi,π µπ (s) = α(s) + γ µi,π (s, a)fi (s, a) s∈S a∈A s∈S a∈A E vi µπ (s , a)T (s , a, s), s ∈S a∈A µπ (s, a) = µπ (s), µπ (s, a) 0 a∈A E where the values vi are found by solving k separate optimization problems (k is the number of E features). [sent-176, score-0.264]

74 For each feature i, vi is the value of the optimal policy in the set ΠE under the reward weights w deﬁned as: wi = 1 and wj = 0, ∀j = i. [sent-177, score-0.736]

75 7 Experimental Results To validate our approach, we experimented on two simulated navigation problems: a gridworld and two racetrack domains, taken from (Boularias & Chaib-draa, 2010). [sent-178, score-0.258]

76 The gridworld is divided into non-overlapping regions, and the reward varies depending on the region in which the agent is located. [sent-188, score-0.335]

77 For each region i, there is a feature fi , where fi (s) indicates whether state s is in region i. [sent-189, score-0.182]

78 The expert’s policy π E corresponds to the optimal deterministic policy found by value iteration. [sent-190, score-0.632]

79 In all our experiments on gridworlds, we used only 10 demonstration trajectories, which is a signiﬁcantly small number compared to other methods ( Neu & Szepesvri (2007) for example). [sent-191, score-0.125]

80 0349 Table 1: Gridworld average reward results Table 1 shows the average reward per step of the learned policy, averaged over 103 independent trials of the same duration as the demonstration trajectories. [sent-229, score-0.725]

81 The second observation is that the values of the policies learned by bootstrapped LPAL were between the values of LPAL with Monte Carlo and the optimal ones. [sent-234, score-0.216]

82 In fact, the policy learned by the bootstrapped LPAL is one that minimizes the difference between the expected frequency of a feature using this policy and the maximal one among all the policies that resemble to the expert’s policy. [sent-235, score-0.854]

83 Therefore, the learned policy maximizes the frequency of a feature that is not necessary a good one (with a high reward weight). [sent-236, score-0.636]

84 Finally, we remark that k-NN performed as an expert in this experiment. [sent-240, score-0.276]

85 The states correspond to the position of the car on the racetrack and its velocity. [sent-244, score-0.277]

86 For racetrack (1), the car always starts from the same initial position, and the duration of each demonstration trajectory is 20 time-steps. [sent-245, score-0.382]

87 For racetrack (2), the car starts at a random position, and the length of each trajectory is 40 time-steps. [sent-246, score-0.237]

88 A high reward is given for reaching the ﬁnish line, a low cost is associated to each movement, and high cost is associated to driving off-road (or hitting an obstacle). [sent-247, score-0.385]

89 Figure 2 (a-f) shows the average reward per step of the learned policies, the average proportion of off-road steps, and the average number of steps before reaching the ﬁnish line, as a function of the number of trajectories in the demonstration. [sent-248, score-0.412]

90 We ﬁrst notice that k-NN performed poorly, this is principally caused by the effect of driving off-road on both the cumulated reward and the velocity of the car. [sent-249, score-0.306]

91 Contrary to the gridworld experiments, MMP with Monte Carlo achieved good performances on racetrack (1). [sent-251, score-0.258]

92 In fact, by ﬁxing the initial state, the demonstration covers most of the reachable states, and the feature frequencies are accurately estimated from the demonstration. [sent-252, score-0.313]

93 On racetrack (2) however, MMP with MC was unable to learn a good policy because all the states were reachable from the initial distribution. [sent-253, score-0.558]

94 Similarly, LPAL with both MC and bootstrapping failed to achieve good results on racetracks (1) and (2). [sent-254, score-0.198]

95 Finally, we notice the nearly optimal performance of the bootstrapped MMP, on both racetracks (1) and (2). [sent-256, score-0.174]

96 8 Conclusion and Future Work The main question of apprenticeship learning is how to generalize the expert’s policy to states that have not been encountered during the demonstration. [sent-257, score-0.564]

97 Inverse Reinforcement Learning (IRL) provides an efﬁcient answer which consists in ﬁrst learning a reward function that explains the observed behavior, and then using it for the generalization. [sent-258, score-0.261]

98 A strong assumption considered in IRL-based al7 20 Expert MMP + MC MMP + Bootstrapping LPAL + MC LPAL + Bootstrapping k−NN 16 14 12 10 8 32 30 28 26 24 22 6 20 4 2 4 6 8 10 Number of trajectories in the demonstration 2 12 0. [sent-259, score-0.201]

99 We also showed that this problem can be solved by modifying the cost function so that the value of the learned policy is compared to the exact value of a generalized expert’s policy. [sent-272, score-0.359]

100 We also provided theoretical insights on the modiﬁed cost function, showing that it admits the expert’s true reward as a locally optimal solution, under mild conditions. [sent-273, score-0.324]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('mmp', 0.445), ('lpal', 0.369), ('policy', 0.295), ('expert', 0.276), ('reward', 0.261), ('vl', 0.252), ('apprenticeship', 0.208), ('racetrack', 0.203), ('cq', 0.164), ('wt', 0.163), ('mc', 0.156), ('bootstrapping', 0.138), ('vi', 0.132), ('demonstration', 0.125), ('frequencies', 0.123), ('se', 0.095), ('bootstrapped', 0.088), ('policies', 0.081), ('trajectories', 0.076), ('irl', 0.076), ('ratliff', 0.069), ('fi', 0.066), ('syed', 0.059), ('boularias', 0.058), ('occupancy', 0.058), ('planning', 0.056), ('gridworld', 0.055), ('actions', 0.051), ('mini', 0.05), ('margin', 0.048), ('demonstrations', 0.048), ('nn', 0.045), ('racetracks', 0.043), ('cost', 0.04), ('states', 0.04), ('max', 0.04), ('ng', 0.04), ('reinforcement', 0.039), ('abbeel', 0.039), ('arg', 0.036), ('car', 0.034), ('russell', 0.033), ('calculated', 0.033), ('obstacles', 0.033), ('carlo', 0.032), ('monte', 0.032), ('frequency', 0.031), ('abdeslam', 0.029), ('fringe', 0.029), ('hitted', 0.029), ('szepesvri', 0.029), ('ow', 0.029), ('returned', 0.028), ('rk', 0.027), ('gridworlds', 0.025), ('state', 0.025), ('driving', 0.025), ('st', 0.025), ('feature', 0.025), ('schapire', 0.024), ('sm', 0.024), ('rewards', 0.024), ('inverse', 0.024), ('learned', 0.024), ('umar', 0.023), ('gw', 0.023), ('neu', 0.023), ('ramachandran', 0.023), ('optimal', 0.023), ('subgradient', 0.022), ('demonstrated', 0.022), ('proposition', 0.021), ('encountered', 0.021), ('action', 0.021), ('transition', 0.021), ('returns', 0.021), ('nish', 0.021), ('equation', 0.021), ('estimated', 0.02), ('duration', 0.02), ('recovering', 0.02), ('reachable', 0.02), ('def', 0.02), ('notice', 0.02), ('consequently', 0.019), ('bootstrap', 0.019), ('hitting', 0.019), ('agent', 0.019), ('deterministic', 0.019), ('features', 0.018), ('amir', 0.018), ('average', 0.017), ('bagnell', 0.017), ('failed', 0.017), ('occurred', 0.017), ('programming', 0.017), ('divergence', 0.016), ('maximal', 0.015), ('cover', 0.015), ('tuple', 0.015)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 1.0 43 nips-2010-Bootstrapping Apprenticeship Learning

Author: Abdeslam Boularias, Brahim Chaib-draa

2 0.35437945 93 nips-2010-Feature Construction for Inverse Reinforcement Learning

Author: Sergey Levine, Zoran Popovic, Vladlen Koltun

Abstract: The goal of inverse reinforcement learning is to ﬁnd a reward function for a Markov decision process, given example traces from its optimal policy. Current IRL techniques generally rely on user-supplied features that form a concise basis for the reward. We present an algorithm that instead constructs reward features from a large collection of component features, by building logical conjunctions of those component features that are relevant to the example policy. Given example traces, the algorithm returns a reward function as well as the constructed features. The reward function can be used to recover a full, deterministic, stationary policy, and the features can be used to transplant the reward function into any novel environment on which the component features are well deﬁned. 1

3 0.35060742 14 nips-2010-A Reduction from Apprenticeship Learning to Classification

Author: Umar Syed, Robert E. Schapire

Abstract: We provide new theoretical results for apprenticeship learning, a variant of reinforcement learning in which the true reward function is unknown, and the goal is to perform well relative to an observed expert. We study a common approach to learning from expert demonstrations: using a classiﬁcation algorithm to learn to imitate the expert’s behavior. Although this straightforward learning strategy is widely-used in practice, it has been subject to very little formal analysis. We prove that, if the learned classiﬁer has error rate ǫ, the difference between the √ value of the apprentice’s policy and the expert’s policy is O( ǫ). Further, we prove that this difference is only O(ǫ) when the expert’s policy is close to optimal. This latter result has an important practical consequence: Not only does imitating a near-optimal expert result in a better policy, but far fewer demonstrations are required to successfully imitate such an expert. This suggests an opportunity for substantial savings whenever the expert is known to be good, but demonstrations are expensive or difﬁcult to obtain. 1

4 0.32786447 184 nips-2010-Nonparametric Bayesian Policy Priors for Reinforcement Learning

Author: Finale Doshi-velez, David Wingate, Nicholas Roy, Joshua B. Tenenbaum

Abstract: We consider reinforcement learning in partially observable domains where the agent can query an expert for demonstrations. Our nonparametric Bayesian approach combines model knowledge, inferred from expert information and independent exploration, with policy knowledge inferred from expert trajectories. We introduce priors that bias the agent towards models with both simple representations and simple policies, resulting in improved policy and model learning. 1

5 0.23594207 189 nips-2010-On a Connection between Importance Sampling and the Likelihood Ratio Policy Gradient

Author: Tang Jie, Pieter Abbeel

Abstract: Likelihood ratio policy gradient methods have been some of the most successful reinforcement learning algorithms, especially for learning on physical systems. We describe how the likelihood ratio policy gradient can be derived from an importance sampling perspective. This derivation highlights how likelihood ratio methods under-use past experience by (i) using the past experience to estimate only the gradient of the expected return U (θ) at the current policy parameterization θ, rather than to obtain a more complete estimate of U (θ), and (ii) using past experience under the current policy only rather than using all past experience to improve the estimates. We present a new policy search method, which leverages both of these observations as well as generalized baselines—a new technique which generalizes commonly used baseline techniques for policy gradient methods. Our algorithm outperforms standard likelihood ratio policy gradient algorithms on several testbeds. 1

6 0.21899027 208 nips-2010-Policy gradients in linearly-solvable MDPs

7 0.21888399 152 nips-2010-Learning from Logged Implicit Exploration Data

8 0.21184239 179 nips-2010-Natural Policy Gradient Methods with Parameter-based Exploration for Control Tasks

9 0.19860351 229 nips-2010-Reward Design via Online Gradient Ascent

10 0.19474012 196 nips-2010-Online Markov Decision Processes under Bandit Feedback

11 0.18007798 160 nips-2010-Linear Complementarity for Regularized Policy Evaluation and Improvement

12 0.1546426 130 nips-2010-Interval Estimation for Reinforcement-Learning Algorithms in Continuous-State Domains

13 0.14545643 78 nips-2010-Error Propagation for Approximate Policy and Value Iteration

14 0.12655425 134 nips-2010-LSTD with Random Projections

15 0.11886188 212 nips-2010-Predictive State Temporal Difference Learning

16 0.10820178 11 nips-2010-A POMDP Extension with Belief-dependent Rewards

17 0.10757613 201 nips-2010-PAC-Bayesian Model Selection for Reinforcement Learning

18 0.1060978 4 nips-2010-A Computational Decision Theory for Interactive Assistants

19 0.093710855 50 nips-2010-Constructing Skill Trees for Reinforcement Learning Agents from Demonstration Trajectories

20 0.09117227 38 nips-2010-Batch Bayesian Optimization via Simulation Matching

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, 0.196), (1, -0.439), (2, -0.066), (3, -0.067), (4, 0.008), (5, -0.03), (6, 0.011), (7, -0.016), (8, 0.018), (9, -0.024), (10, -0.045), (11, 0.054), (12, 0.043), (13, -0.008), (14, 0.015), (15, -0.042), (16, -0.028), (17, 0.038), (18, -0.007), (19, -0.055), (20, 0.017), (21, -0.077), (22, 0.016), (23, -0.04), (24, 0.074), (25, 0.011), (26, -0.026), (27, 0.041), (28, 0.042), (29, -0.053), (30, -0.034), (31, 0.059), (32, 0.004), (33, 0.037), (34, -0.078), (35, -0.064), (36, 0.063), (37, 0.041), (38, 0.22), (39, -0.098), (40, -0.065), (41, 0.081), (42, 0.133), (43, 0.036), (44, -0.087), (45, 0.018), (46, 0.037), (47, 0.048), (48, 0.017), (49, 0.059)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.96443027 43 nips-2010-Bootstrapping Apprenticeship Learning

Author: Abdeslam Boularias, Brahim Chaib-draa

2 0.83810228 93 nips-2010-Feature Construction for Inverse Reinforcement Learning

Author: Sergey Levine, Zoran Popovic, Vladlen Koltun

3 0.80598587 184 nips-2010-Nonparametric Bayesian Policy Priors for Reinforcement Learning

Author: Finale Doshi-velez, David Wingate, Nicholas Roy, Joshua B. Tenenbaum

4 0.78362072 14 nips-2010-A Reduction from Apprenticeship Learning to Classification

Author: Umar Syed, Robert E. Schapire

5 0.71967608 229 nips-2010-Reward Design via Online Gradient Ascent

Author: Jonathan Sorg, Richard L. Lewis, Satinder P. Singh

Abstract: Recent work has demonstrated that when artiﬁcial agents are limited in their ability to achieve their goals, the agent designer can beneﬁt by making the agent’s goals different from the designer’s. This gives rise to the optimization problem of designing the artiﬁcial agent’s goals—in the RL framework, designing the agent’s reward function. Existing attempts at solving this optimal reward problem do not leverage experience gained online during the agent’s lifetime nor do they take advantage of knowledge about the agent’s structure. In this work, we develop a gradient ascent approach with formal convergence guarantees for approximately solving the optimal reward problem online during an agent’s lifetime. We show that our method generalizes a standard policy gradient approach, and we demonstrate its ability to improve reward functions in agents with various forms of limitations. 1 The Optimal Reward Problem In this work, we consider the scenario of an agent designer building an autonomous agent. The designer has his or her own goals which must be translated into goals for the autonomous agent. We represent goals using the Reinforcement Learning (RL) formalism of the reward function. This leads to the optimal reward problem of designing the agent’s reward function so as to maximize the objective reward received by the agent designer. Typically, the designer assigns his or her own reward to the agent. However, there is ample work which demonstrates the beneﬁt of assigning reward which does not match the designer’s. For example, work on reward shaping [11] has shown how to modify rewards to accelerate learning without altering the optimal policy, and PAC-MDP methods [5, 20] including approximate Bayesian methods [7, 19] add bonuses to the objective reward to achieve optimism under uncertainty. These approaches explicitly or implicitly assume that the asymptotic behavior of the agent should be the same as that which would occur using the objective reward function. These methods do not explicitly consider the optimal reward problem; however, they do show improved performance through reward modiﬁcation. In our recent work that does explicitly consider the optimal reward problem [18], we analyzed an explicit hypothesis about the beneﬁt of reward design—that it helps mitigate the performance loss caused by computational constraints (bounds) on agent architectures. We considered various types of agent limitations—limits on planning depth, failure to account for partial observability, and other erroneous modeling assumptions—and demonstrated the beneﬁts of good reward functions in each case empirically. Crucially, in bounded agents, the optimal reward function often leads to behavior that is different from the asymptotic behavior achieved with the objective reward function. In this work, we develop an algorithm, Policy Gradient for Reward Design (PGRD), for improving reward functions for a family of bounded agents that behave according to repeated local (from the current state) model-based planning. We show that this algorithm is capable of improving the reward functions in agents with computational limitations necessitating small bounds on the depth of planning, and also from the use of an inaccurate model (which may be inaccurate due to computationally-motivated approximations). PGRD has few parameters, improves the reward 1 function online during an agent’s lifetime, takes advantage of knowledge about the agent’s structure (through the gradient computation), and is linear in the number of reward function parameters. Notation. Formally, we consider discrete-time partially-observable environments with a ﬁnite number of hidden states s ∈ S, actions a ∈ A, and observations o ∈ O; these ﬁnite set assumptions are useful for our theorems, but our algorithm can handle inﬁnite sets in practice. Its dynamics are governed by a state-transition function P (s |s, a) that deﬁnes a distribution over next-states s conditioned on current state s and action a, and an observation function Ω(o|s) that deﬁnes a distribution over observations o conditioned on current state s. The agent designer’s goals are speciﬁed via the objective reward function RO . At each time step, the designer receives reward RO (st ) ∈ [0, 1] based on the current state st of the environment, where the subscript denotes time. The designer’s objective return is the expected mean objective reward N 1 obtained over an inﬁnite horizon, i.e., limN →∞ E N t=0 RO (st ) . In the standard view of RL, the agent uses the same reward function as the designer to align the interests of the agent and the designer. Here we allow for a separate agent reward function R(· ). An agent’s reward function can in general be deﬁned in terms of the history of actions and observations, but is often more pragmatically deﬁned in terms of some abstraction of history. We deﬁne the agent’s reward function precisely in Section 2. Optimal Reward Problem. An RL agent attempts to act so as to maximize its own cumulative reward, or return. Crucially, as a result, the sequence of environment-states {st }∞ is affected by t=0 the choice of reward function; therefore, the agent designer’s return is affected as well. The optimal reward problem arises from the fact that while the objective reward function is ﬁxed as part of the problem description, the reward function is a choice to be made by the designer. We capture this choice abstractly by letting the reward be parameterized by some vector of parameters θ chosen from space of parameters Θ. Each θ ∈ Θ speciﬁes a reward function R(· ; θ) which in turn produces a distribution over environment state sequences via whatever RL method the agent uses. The expected N 1 return obtained by the designer for choice θ is U(θ) = limN →∞ E N t=0 RO (st ) R(·; θ) . The optimal reward parameters are given by the solution to the optimal reward problem [16, 17, 18]: θ∗ = arg max U(θ) = arg max lim E θ∈Θ θ∈Θ N →∞ 1 N N RO (st ) R(·; θ) . (1) t=0 Our previous research on solving the optimal reward problem has focused primarily on the properties of the optimal reward function and its correspondence to the agent architecture and the environment [16, 17, 18]. This work has used inefﬁcient exhaustive search methods for ﬁnding good approximations to θ∗ (though there is recent work on using genetic algorithms to do this [6, 9, 12]). Our primary contribution in this paper is a new convergent online stochastic gradient method for ﬁnding approximately optimal reward functions. To our knowledge, this is the ﬁrst algorithm that improves reward functions in an online setting—during a single agent’s lifetime. In Section 2, we present the PGRD algorithm, prove its convergence, and relate it to OLPOMDP [2], a policy gradient algorithm. In Section 3, we present experiments demonstrating PGRD’s ability to approximately solve the optimal reward problem online. 2 PGRD: Policy Gradient for Reward Design PGRD builds on the following insight: the agent’s planning algorithm procedurally converts the reward function into behavior; thus, the reward function can be viewed as a speciﬁc parameterization of the agent’s policy. Using this insight, PGRD updates the reward parameters by estimating the gradient of the objective return with respect to the reward parameters, θ U(θ), from experience, using standard policy gradient techniques. In fact, we show that PGRD can be viewed as an (independently interesting) generalization of the policy gradient method OLPOMDP [2]. Speciﬁcally, we show that OLPOMDP is special case of PGRD when the planning depth d is zero. In this section, we ﬁrst present the family of local planning agents for which PGRD improves the reward function. Next, we develop PGRD and prove its convergence. Finally, we show that PGRD generalizes OLPOMDP and discuss how adding planning to OLPOMDP affects the space of policies available to the optimization method. 2 1 2 3 4 5 Input: T , θ0 , {αt }∞ , β, γ t=0 o0 , i0 = initializeStart(); for t = 0, 1, 2, 3, . . . do ∀a Qt (a; θt ) = plan(it , ot , T, R(it , ·, ·; θt ), d,γ); at ∼ µ(a|it ; Qt ); rt+1 , ot+1 = takeAction(at ); µ(a |i ;Q ) 6 7 8 9 t zt+1 = βzt + θt t |itt ;Qt ) t ; µ(a θt+1 = θt + αt (rt+1 zt+1 − λθt ) ; it+1 = updateInternalState(it , at , ot+1 ); end Figure 1: PGRD (Policy Gradient for Reward Design) Algorithm A Family of Limited Agents with Internal State. Given a Markov model T deﬁned over the observation space O and action space A, denote T (o |o, a) the probability of next observation o given that the agent takes action a after observing o. Our agents use the model T to plan. We do not assume that the model T is an accurate model of the environment. The use of an incorrect model is one type of agent limitation we examine in our experiments. In general, agents can use non-Markov models deﬁned in terms of the history of observations and actions; we leave this for future work. The agent maintains an internal state feature vector it that is updated at each time step using it+1 = updateInternalState(it , at , ot+1 ). The internal state allows the agent to use reward functions T that depend on the agent’s history. We consider rewards of the form R(it , o, a; θt ) = θt φ(it , o, a), where θt is the reward parameter vector at time t, and φ(it , o, a) is a vector of features based on internal state it , planning state o, and action a. Note that if φ is a vector of binary indicator features, this representation allows for arbitrary reward functions and thus the representation is completely general. Many existing methods use reward functions that depend on history. Reward functions based on empirical counts of observations, as in PAC-MDP approaches [5, 20], provide some examples; see [14, 15, 13] for others. We present a concrete example in our empirical section. At each time step t, the agent’s planning algorithm, plan, performs depth-d planning using the model T and reward function R(it , o, a; θt ) with current internal state it and reward parameters θt . Speciﬁcally, the agent computes a d-step Q-value function Qd (it , ot , a; θt ) ∀a ∈ A, where Qd (it , o, a; θt ) = R(it , o, a; θt ) + γ o ∈O T (o |o, a) maxb∈A Qd−1 (it , o , b; θt ) and Q0 (it , o, a; θt ) = R(it , o, a; θt ). We emphasize that the internal state it and reward parameters θt are held invariant while planning. Note that the d-step Q-values are only computed for the current observation ot , in effect by building a depth-d tree rooted at ot . In the d = 0 special case, the planning procedure completely ignores the model T and returns Q0 (it , ot , a; θt ) = R(it , ot , a; θt ). Regardless of the value of d, we treat the end result of planning as providing a scoring function Qt (a; θt ) where the dependence on d, it and ot is dropped from the notation. To allow for gradient calculations, our agents act according to the τ Qt (a;θt ) def Boltzmann (soft-max) stochastic policy parameterized by Q: µ(a|it ; Qt ) = e eτ Qt (b;θt ) , where τ b is a temperature parameter that determines how stochastically the agent selects the action with the highest score. When the planning depth d is small due to computational limitations, the agent cannot account for events beyond the planning depth. We examine this limitation in our experiments. Gradient Ascent. To develop a gradient algorithm for improving the reward function, we need to compute the gradient of the objective return with respect to θ: θ U(θ). The main insight is to break the gradient calculation into the calculation of two gradients. The ﬁrst is the gradient of the objective return with respect to the policy µ, and the second is the gradient of the policy with respect to the reward function parameters θ. The ﬁrst gradient is exactly what is computed in standard policy gradient approaches [2]. The second gradient is challenging because the transformation from reward parameters to policy involves a model-based planning procedure. We draw from the work of Neu and Szepesv´ ri [10] which shows that this gradient computation resembles planning itself. We a develop PGRD, presented in Figure 1, explicitly as a generalization of OLPOMDP, a policy gradient algorithm developed by Bartlett and Baxter [2], because of its foundational simplicity relative to other policy-gradient algorithms such as those based on actor-critic methods (e.g., [4]). Notably, the reward parameters are the only parameters being learned in PGRD. 3 PGRD follows the form of OLPOMDP (Algorithm 1 in Bartlett and Baxter [2]) but generalizes it in three places. In Figure 1 line 3, the agent plans to compute the policy, rather than storing the policy directly. In line 6, the gradient of the policy with respect to the parameters accounts for the planning procedure. In line 8, the agent maintains a general notion of internal state that allows for richer parameterization of policies than typically considered (similar to Aberdeen and Baxter [1]). The algorithm takes as parameters a sequence of learning rates {αk }, a decaying-average parameter β, and regularization parameter λ > 0 which keeps the the reward parameters θ bounded throughout learning. Given a sequence of calculations of the gradient of the policy with respect to the parameters, θt µ(at |it ; Qt ), the remainder of the algorithm climbs the gradient of objective return θ U(θ) using OLPOMDP machinery. In the next subsection, we discuss how to compute θt µ(at |it ; Qt ). Computing the Gradient of the Policy with respect to Reward. For the Boltzmann distribution, the gradient of the policy with respect to the reward parameters is given by the equation θt µ(a|it ; Qt ) = τ · µ(a|Qt )[ θt Qt (a|it ; θt ) − θt Qt (b; θt )], where τ is the Boltzmann b∈A temperature (see [10]). Thus, computing θt µ(a|it ; Qt ) reduces to computing θt Qt (a; θt ). The value of Qt depends on the reward parameters θt , the model, and the planning depth. However, as we present below, the process of computing the gradient closely resembles the process of planning itself, and the two computations can be interleaved. Theorem 1 presented below is an adaptation of Proposition 4 from Neu and Szepesv´ ri [10]. It presents the gradient computation for depth-d a planning as well as for inﬁnite-depth discounted planning. We assume that the gradient of the reward function with respect to the parameters is bounded: supθ,o,i,a θ R(i, o, a, θ) < ∞. The proof of the theorem follows directly from Proposition 4 of Neu and Szepesv´ ri [10]. a Theorem 1. Except on a set of measure zero, for any depth d, the gradient θ Qd (o, a; θ) exists and is given by the recursion (where we have dropped the dependence on i for simplicity) d θ Q (o, a; θ) = θ R(o, a; θ) π d−1 (b|o ) T (o |o, a) +γ o ∈O d−1 (o θQ , b; θ), (2) b∈A where θ Q0 (o, a; θ) = θ R(o, a; θ) and π d (a|o) ∈ arg maxa Qd (o, a; θ) is any policy that is greedy with respect to Qd . The result also holds for θ Q∗ (o, a; θ) = θ limd→∞ Qd (o, a; θ). The Q-function will not be differentiable when there are multiple optimal policies. This is reﬂected in the arbitrary choice of π in the gradient calculation. However, it was shown by Neu and Szepesv´ ri [10] that even for values of θ which are not differentiable, the above computation produces a a valid calculation of a subgradient; we discuss this below in our proof of convergence of PGRD. Convergence of PGRD (Figure 1). Given a particular ﬁxed reward function R(·; θ), transition model T , and planning depth, there is a corresponding ﬁxed randomized policy µ(a|i; θ)—where we have explicitly represented the reward’s dependence on the internal state vector i in the policy parameterization and dropped Q from the notation as it is redundant given that everything else is ﬁxed. Denote the agent’s internal-state update as a (usually deterministic) distribution ψ(i |i, a, o). Given a ﬁxed reward parameter vector θ, the joint environment-state–internal-state transitions can be modeled as a Markov chain with a |S||I| × |S||I| transition matrix M (θ) whose entries are given by M s,i , s ,i (θ) = p( s , i | s, i ; θ) = o,a ψ(i |i, a, o)Ω(o|s )P (s |s, a)µ(a|i; θ). We make the following assumptions about the agent and the environment: Assumption 1. The transition matrix M (θ) of the joint environment-state–internal-state Markov chain has a unique stationary distribution π(θ) = [πs1 ,i1 (θ), πs2 ,i2 (θ), . . . , πs|S| ,i|I| (θ)] satisfying the balance equations π(θ)M (θ) = π(θ), for all θ ∈ Θ. Assumption 2. During its execution, PGRD (Figure 1) does not reach a value of it , and θt at which µ(at |it , Qt ) is not differentiable with respect to θt . It follows from Assumption 1 that the objective return, U(θ), is independent of the start state. The original OLPOMDP convergence proof [2] has a similar condition that only considers environment states. Intuitively, this condition allows PGRD to handle history-dependence of a reward function in the same manner that it handles partial observability in an environment. Assumption 2 accounts for the fact that a planning algorithm may not be fully differentiable everywhere. However, Theorem 1 showed that inﬁnite and bounded-depth planning is differentiable almost everywhere (in a measure theoretic sense). Furthermore, this assumption is perhaps stronger than necessary, as stochastic approximation algorithms, which provide the theory upon which OLPOMDP is based, have been shown to converge using subgradients [8]. 4 In order to state the convergence theorem, we must deﬁne the approximate gradient which OLPOMDP def T calculates. Let the approximate gradient estimate be β U(θ) = limT →∞ t=1 rt zt for a ﬁxed θ and θ PGRD parameter β, where zt (in Figure 1) represents a time-decaying average of the θt µ(at |it , Qt ) calculations. It was shown by Bartlett and Baxter [2] that β U(θ) is close to the true value θ U(θ) θ for large values of β. Theorem 2 proves that PGRD converges to a stable equilibrium point based on this approximate gradient measure. This equilibrium point will typically correspond to some local optimum in the return function U(θ). Given our development and assumptions, the theorem is a straightforward extension of Theorem 6 from Bartlett and Baxter [2] (proof omitted). ∞ Theorem 2. Given β ∈ [0, 1), λ > 0, and a sequence of step sizes αt satisfying t=0 αt = ∞ and ∞ 2 t=0 (αt ) < ∞, PGRD produces a sequence of reward parameters θt such that θt → L as t → ∞ a.s., where L is the set of stable equilibrium points of the differential equation ∂θ = β U(θ) − λθ. θ ∂t PGRD generalizes OLPOMDP. As stated above, OLPOMDP, when it uses a Boltzmann distribution in its policy representation (a common case), is a special case of PGRD when the planning depth is zero. First, notice that in the case of depth-0 planning, Q0 (i, o, a; θ) = R(i, o, a, θ), regardless of the transition model and reward parameterization. We can also see from Theorem 1 that 0 θ Q (i, o, a; θ) = θ R(i, o, a; θ). Because R(i, o, a; θ) can be parameterized arbitrarily, PGRD can be conﬁgured to match standard OLPOMDP with any policy parameterization that also computes a score function for the Boltzmann distribution. In our experiments, we demonstrate that choosing a planning depth d > 0 can be beneﬁcial over using OLPOMDP (d = 0). In the remainder of this section, we show theoretically that choosing d > 0 does not hurt in the sense that it does not reduce the space of policies available to the policy gradient method. Speciﬁcally, we show that when using an expressive enough reward parameterization, PGRD’s space of policies is not restricted relative to OLPOMDP’s space of policies. We prove the result for inﬁnite planning, but the extension to depth-limited planning is straightforward. Theorem 3. There exists a reward parameterization such that, for an arbitrary transition model T , the space of policies representable by PGRD with inﬁnite planning is identical to the space of policies representable by PGRD with depth 0 planning. Proof. Ignoring internal state for now (holding it constant), let C(o, a) be an arbitrary reward function used by PGRD with depth 0 planning. Let R(o, a; θ) be a reward function for PGRD with inﬁnite (d = ∞) planning. The depth-∞ agent uses the planning result Q∗ (o, a; θ) to act, while the depth-0 agent uses the function C(o, a) to act. Therefore, it sufﬁces to show that one can always choose θ such that the planning solution Q∗ (o, a; θ) equals C(o, a). For all o ∈ O, a ∈ A, set R(o, a; θ) = C(o, a) − γ o T (o |o, a) maxa C(o , a ). Substituting Q∗ for C, this is the Bellman optimality equation [22] for inﬁnite-horizon planning. Setting R(o, a; θ) as above is possible if it is parameterized by a table with an entry for each observation–action pair. Theorem 3 also shows that the effect of an arbitrarily poor model can be overcome with a good choice of reward function. This is because a Boltzmann distribution can, allowing for an arbitrary scoring function C, represent any policy. We demonstrate this ability of PGRD in our experiments. 3 Experiments The primary objective of our experiments is to demonstrate that PGRD is able to use experience online to improve the reward function parameters, thereby improving the agent’s obtained objective return. Speciﬁcally, we compare the objective return achieved by PGRD to the objective return achieved by PGRD with the reward adaptation turned off. In both cases, the reward function is initialized to the objective reward function. A secondary objective is to demonstrate that when a good model is available, adding the ability to plan—even for small depths—improves performance relative to the baseline algorithm of OLPOMDP (or equivalently PGRD with depth d = 0). Foraging Domain for Experiments 1 to 3: The foraging environment illustrated in Figure 2(a) is a 3 × 3 grid world with 3 dead-end corridors (rows) separated by impassable walls. The agent (bird) has four available actions corresponding to each cardinal direction. Movement in the intended direction fails with probability 0.1, resulting in movement in a random direction. If the resulting direction is 5 Objective Return 0.15 D=6, α=0 & D=6, α=5×10 −5 D=4, α=2×10 −4 D=0, α=5×10 −4 0.1 0.05 0 D=4, α=0 D=0, α=0 1000 2000 3000 4000 5000 Time Steps C) Objective Return B) A) 0.15 D=6, α=0 & D=6, α=5×10 −5 D=3, α=3×10 −3 D=1, α=3×10 −4 0.1 D=3, α=0 0.05 D=0, α=0.01 & D=1, α=0 0 1000 2000 3000 4000 5000 D=0, α=0 Time Steps Figure 2: A) Foraging Domain, B) Performance of PGRD with observation-action reward features, C) Performance of PGRD with recency reward features blocked by a wall or the boundary, the action results in no movement. There is a food source (worm) located in one of the three right-most locations at the end of each corridor. The agent has an eat action, which consumes the worm when the agent is at the worm’s location. After the agent consumes the worm, a new worm appears randomly in one of the other two potential worm locations. Objective Reward for the Foraging Domain: The designer’s goal is to maximize the average number of worms eaten per time step. Thus, the objective reward function RO provides a reward of 1.0 when the agent eats a worm, and a reward of 0 otherwise. The objective return is deﬁned as in Equation (1). Experimental Methodology: We tested PGRD for depth-limited planning agents of depths 0–6. Recall that PGRD for the agent with planning depth 0 is the OLPOMDP algorithm. For each depth, we jointly optimized over the PGRD algorithm parameters, α and β (we use a ﬁxed α throughout learning). We tested values for α on an approximate logarithmic scale in the range (10−6 , 10−2 ) as well as the special value of α = 0, which corresponds to an agent that does not adapt its reward function. We tested β values in the set 0, 0.4, 0.7, 0.9, 0.95, 0.99. Following common practice [3], we set the λ parameter to 0. We explicitly bound the reward parameters and capped the reward function output both to the range [−1, 1]. We used a Boltzmann temperature parameter of τ = 100 and planning discount factor γ = 0.95. Because we initialized θ so that the initial reward function was the objective reward function, PGRD with α = 0 was equivalent to a standard depth-limited planning agent. Experiment 1: A fully observable environment with a correct model learned online. In this experiment, we improve the reward function in an agent whose only limitation is planning depth, using (1) a general reward parameterization based on the current observation and (2) a more compact reward parameterization which also depends on the history of observations. Observation: The agent observes the full state, which is given by the pair o = (l, w), where l is the agent’s location and w is the worm’s location. Learning a Correct Model: Although the theorem of convergence of PGRD relies on the agent having a ﬁxed model, the algorithm itself is readily applied to the case of learning a model online. In this experiment, the agent’s model T is learned online based on empirical transition probabilities between observations (recall this is a fully observable environment). Let no,a,o be the number of times that o was reached after taking action a after observing o. The agent models the probability of seeing o as no,a,o T (o |o, a) = . n o o,a,o Reward Parameterizations: Recall that R(i, o, a; θ) = θT φ(i, o, a), for some φ(i, o, a). (1) In the observation-action parameterization, φ(i, o, a) is a binary feature vector with one binary feature for each observation-action pair—internal state is ignored. This is effectively a table representation over all reward functions indexed by (o, a). As shown in Theorem 3, the observation-action feature representation is capable of producing arbitrary policies over the observations. In large problems, such a parameterization would not be feasible. (2) The recency parameterization is a more compact representation which uses features that rely on the history of observations. The feature vector is φ(i, o, a) = [RO (o, a), 1, φcl (l, i), φcl,a (l, a, i)], where RO (o, a) is the objective reward function deﬁned as above. The feature φcl (l) = 1 − 1/c(l, i), where c(l, i) is the number of time steps since the agent has visited location l, as represented in the agent’s internal state i. Its value is normalized to the range [0, 1) and is high when the agent has not been to location l recently. The feature φcl,a (l, a, i) = 1 − 1/c(l, a, i) is similarly deﬁned with respect to the time since the agent has taken action a in location l. Features based on recency counts encourage persistent exploration [21, 18]. 6 Results & Discussion: Figure 2(b) and Figure 2(c) present results for agents that use the observationaction parameterization and the recency parameterization of the reward function respectively. The horizontal axis is the number of time steps of experience. The vertical axis is the objective return, i.e., the average objective reward per time step. Each curve is an average over 130 trials. The values of d and the associated optimal algorithm parameters for each curve are noted in the ﬁgures. First, note that with d = 6, the agent is unbounded, because food is never more than 6 steps away. Therefore, the agent does not beneﬁt from adapting the reward function parameters (given that we initialize to the objective reward function). Indeed, the d = 6, α = 0 agent performs as well as the best reward-optimizing agent. The performance for d = 6 improves with experience because the model improves with experience (and thus from the curves it is seen that the model gets quite accurate in about 1500 time steps). The largest objective return obtained for d = 6 is also the best objective return that can be obtained for any value of d. Several results can be observed in both Figures 2(b) and (c). 1) Each curve that uses α > 0 (solid lines) improves with experience. This is a demonstration of our primary contribution, that PGRD is able to effectively improve the reward function with experience. That the improvement over time is not just due to model learning is seen in the fact that for each value of d < 6 the curve for α > 0 (solid-line) which adapts the reward parameters does signiﬁcantly better than the corresponding curve for α = 0 (dashed-line); the α = 0 agents still learn the model. 2) For both α = 0 and α > 0 agents, the objective return obtained by agents with equivalent amounts of experience increases monotonically as d is increased (though to maintain readability we only show selected values of d in each ﬁgure). This demonstrates our secondary contribution, that the ability to plan in PGRD signiﬁcantly improves performance over standard OLPOMDP (PGRD with d = 0). There are also some interesting differences between the results for the two different reward function parameterizations. With the observation-action parameterization, we noted that there always exists a setting of θ for all d that will yield optimal objective return. This is seen in Figure 2(b) in that all solid-line curves approach optimal objective return. In contrast, the more compact recency reward parameterization does not afford this guarantee and indeed for small values of d (< 3), the solid-line curves in Figure 2(c) converge to less than optimal objective return. Notably, OLPOMDP (d = 0) does not perform well with this feature set. On the other hand, for planning depths 3 ≤ d < 6, the PGRD agents with the recency parameterization achieve optimal objective return faster than the corresponding PGRD agent with the observation-action parameterization. Finally, we note that this experiment validates our claim that PGRD can improve reward functions that depend on history. Experiment 2: A fully observable environment and poor given model. Our theoretical analysis showed that PGRD with an incorrect model and the observation–action reward parameterization should (modulo local maxima issues) do just as well asymptotically as it would with a correct model. Here we illustrate this theoretical result empirically on the same foraging domain and objective reward function used in Experiment 1. We also test our hypothesis that a poor model should slow down the rate of learning relative to a correct model. Poor Model: We gave the agents a ﬁxed incorrect model of the foraging environment that assumes there are no internal walls separating the 3 corridors. Reward Parameterization: We used the observation–action reward parameterization. With a poor model it is no longer interesting to initialize θ so that the initial reward function is the objective reward function because even for d = 6 such an agent would do poorly. Furthermore, we found that this initialization leads to excessively bad exploration and therefore poor learning of how to modify the reward. Thus, we initialize θ to uniform random values near 0, in the range (−10−3 , 10−3 ). Results: Figure 3(a) plots the objective return as a function of number of steps of experience. Each curve is an average over 36 trials. As hypothesized, the bad model slows learning by a factor of more than 10 (notice the difference in the x-axis scales from those in Figure 2). Here, deeper planning results in slower learning and indeed the d = 0 agent that does not use the model at all learns the fastest. However, also as hypothesized, because they used the expressive observation–action parameterization, agents of all planning depths mitigate the damage caused by the poor model and eventually converge to the optimal objective return. Experiment 3: Partially observable foraging world. Here we evaluate PGRD’s ability to learn in a partially observable version of the foraging domain. In addition, the agents learn a model under the erroneous (and computationally convenient) assumption that the domain is fully observable. 7 0.1 −4 D = 0, α = 2 ×10 D = 2, α = 3 ×10 −5 −5 D = 6, α = 2 ×10 0.05 D = 0&2&6, α = 0 0 1 2 3 Time Steps 4 5 x 10 4 0.06 D = 6, α = 7 ×10 D = 2, α = 7 ×10 −4 0.04 D = 1, α = 7 ×10 −4 D = 0, α = 5 ×10 −4 D = 0, α = 0 D = 1&2&6, α = 0 0.02 0 C) −4 1000 2000 3000 4000 5000 Time Steps Objective Return B) 0.08 0.15 Objective Return Objective Return A) 2.5 2 x 10 −3 D=6, α=3×10 −6 D=0, α=1×10 −5 1.5 D=0&6, α=0 1 0.5 1 2 3 Time Steps 4 5 x 10 4 Figure 3: A) Performance of PGRD with a poor model, B) Performance of PGRD in a partially observable world with recency reward features, C) Performance of PGRD in Acrobot Partial Observation: Instead of viewing the location of the worm at all times, the agent can now only see the worm when it is colocated with it: its observation is o = (l, f ), where f indicates whether the agent is colocated with the food. Learning an Incorrect Model: The model is learned just as in Experiment 1. Because of the erroneous full observability assumption, the model will hallucinate about worms at all the corridor ends based on the empirical frequency of having encountered them there. Reward Parameterization: We used the recency parameterization; due to the partial observability, agents with the observation–action feature set perform poorly in this environment. The parameters θ are initialized such that the initial reward function equals the objective reward function. Results & Discussion: Figure 3(b) plots the mean of 260 trials. As seen in the solid-line curves, PGRD improves the objective return at all depths (only a small amount for d = 0 and signiﬁcantly more for d > 0). In fact, agents which don’t adapt the reward are hurt by planning (relative to d = 0). This experiment demonstrates that the combination of planning and reward improvement can be beneﬁcial even when the model is erroneous. Because of the partial observability, optimal behavior in this environment achieves less objective return than in Experiment 1. Experiment 4: Acrobot. In this experiment we test PGRD in the Acrobot environment [22], a common benchmark task in the RL literature and one that has previously been used in the testing of policy gradient approaches [23]. This experiment demonstrates PGRD in an environment in which an agent must be limited due to the size of the state space and further demonstrates that adding model-based planning to policy gradient approaches can improve performance. Domain: The version of Acrobot we use is as speciﬁed by Sutton and Barto [22]. It is a two-link robot arm in which the position of one shoulder-joint is ﬁxed and the agent’s control is limited to 3 actions which apply torque to the elbow-joint. Observation: The fully-observable state space is 4 dimensional, with two joint angles ψ1 and ψ2 , and ˙ ˙ two joint velocities ψ1 and ψ2 . Objective Reward: The designer receives an objective reward of 1.0 when the tip is one arm’s length above the ﬁxed shoulder-joint, after which the bot is reset to its initial resting position. Model: We provide the agent with a perfect model of the environment. Because the environment is continuous, value iteration is intractable, and computational limitations prevent planning deep enough to compute the optimal action in any state. The feature vector contains 13 entries. One feature corresponds to the objective reward signal. For each action, there are 5 features corresponding to each of the state features plus an additional feature representing the height of the tip: φ(i, o, a) = ˙ ˙ [RO (o), {ψ1 (o), ψ2 (o), ψ1 (o), ψ2 (o), h(o)}a ]. The height feature has been used in previous work as an alternative deﬁnition of objective reward [23]. Results & Discussion: We plot the mean of 80 trials in Figure 3(c). Agents that use the ﬁxed (α = 0) objective reward function with bounded-depth planning perform according to the bottom two curves. Allowing PGRD and OLPOMDP to adapt the parameters θ leads to improved objective return, as seen in the top two curves in Figure 3(c). Finally, the PGRD d = 6 agent outperforms the standard OLPOMDP agent (PGRD with d = 0), further demonstrating that PGRD outperforms OLPOMDP. Overall Conclusion: We developed PGRD, a new method for approximately solving the optimal reward problem in bounded planning agents that can be applied in an online setting. We showed that PGRD is a generalization of OLPOMDP and demonstrated that it both improves reward functions in limited agents and outperforms the model-free OLPOMDP approach. 8 References [1] Douglas Aberdeen and Jonathan Baxter. Scalable Internal-State Policy-Gradient Methods for POMDPs. Proceedings of the Nineteenth International Conference on Machine Learning, 2002. [2] Peter L. Bartlett and Jonathan Baxter. Stochastic optimization of controlled partially observable Markov decision processes. In Proceedings of the 39th IEEE Conference on Decision and Control, 2000. [3] Jonathan Baxter, Peter L. Bartlett, and Lex Weaver. Experiments with Inﬁnite-Horizon, Policy-Gradient Estimation, 2001. [4] Shalabh Bhatnagar, Richard S. Sutton, M Ghavamzadeh, and Mark Lee. Natural actor-critic algorithms. Automatica, 2009. [5] Ronen I. Brafman and Moshe Tennenholtz. R-MAX - A General Polynomial Time Algorithm for NearOptimal Reinforcement Learning. Journal of Machine Learning Research, 3:213–231, 2001. [6] S. Elfwing, Eiji Uchibe, K. Doya, and H. I. Christensen. Co-evolution of Shaping Rewards and MetaParameters in Reinforcement Learning. Adaptive Behavior, 16(6):400–412, 2008. [7] J. Zico Kolter and Andrew Y. Ng. Near-Bayesian exploration in polynomial time. In Proceedings of the 26th International Conference on Machine Learning, pages 513–520, 2009. [8] Harold J. Kushner and G. George Yin. Stochastic Approximation and Recursive Algorithms and Applications. Springer, 2nd edition, 2010. [9] Cetin Mericli, Tekin Mericli, and H. Levent Akin. A Reward Function Generation Method Using Genetic ¸ ¸ ¸ Algorithms : A Robot Soccer Case Study (Extended Abstract). In Proc. of the 9th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2010), number 2, pages 1513–1514, 2010. [10] Gergely Neu and Csaba Szepesv´ ri. Apprenticeship learning using inverse reinforcement learning and a gradient methods. In Proceedings of the 23rd Conference on Uncertainty in Artiﬁcial Intelligence, pages 295–302, 2007. [11] Andrew Y. Ng, Stuart J. Russell, and D. Harada. Policy invariance under reward transformations: Theory and application to reward shaping. In Proceedings of the 16th International Conference on Machine Learning, pages 278–287, 1999. [12] Scott Niekum, Andrew G. Barto, and Lee Spector. Genetic Programming for Reward Function Search. IEEE Transactions on Autonomous Mental Development, 2(2):83–90, 2010. [13] Pierre-Yves Oudeyer, Frederic Kaplan, and Verena V. Hafner. Intrinsic Motivation Systems for Autonomous Mental Development. IEEE Transactions on Evolutionary Computation, 11(2):265–286, April 2007. [14] J¨ rgen Schmidhuber. Curious model-building control systems. In IEEE International Joint Conference on u Neural Networks, pages 1458–1463, 1991. [15] Satinder Singh, Andrew G. Barto, and Nuttapong Chentanez. Intrinsically Motivated Reinforcement Learning. In Proceedings of Advances in Neural Information Processing Systems 17 (NIPS), pages 1281–1288, 2005. [16] Satinder Singh, Richard L. Lewis, and Andrew G. Barto. Where Do Rewards Come From? In Proceedings of the Annual Conference of the Cognitive Science Society, pages 2601–2606, 2009. [17] Satinder Singh, Richard L. Lewis, Andrew G. Barto, and Jonathan Sorg. Intrinsically Motivated Reinforcement Learning: An Evolutionary Perspective. IEEE Transations on Autonomous Mental Development, 2(2):70–82, 2010. [18] Jonathan Sorg, Satinder Singh, and Richard L. Lewis. Internal Rewards Mitigate Agent Boundedness. In Proceedings of the 27th International Conference on Machine Learning, 2010. [19] Jonathan Sorg, Satinder Singh, and Richard L. Lewis. Variance-Based Rewards for Approximate Bayesian Reinforcement Learning. In Proceedings of the 26th Conference on Uncertainty in Artiﬁcial Intelligence, 2010. [20] Alexander L. Strehl and Michael L. Littman. An analysis of model-based Interval Estimation for Markov Decision Processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008. [21] Richard S. Sutton. Integrated Architectures for Learning, Planning, and Reacting Based on Approximating Dynamic Programming. In The Seventh International Conference on Machine Learning, pages 216–224. 1990. [22] Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. The MIT Press, 1998. [23] Lex Weaver and Nigel Tao. The Optimal Reward Baseline for Gradient-Based Reinforcement Learning. In Proceedings of the 17th Conference on Uncertainty in Artiﬁcial Intelligence, pages 538–545. 2001. 9

6 0.6714744 50 nips-2010-Constructing Skill Trees for Reinforcement Learning Agents from Demonstration Trajectories

7 0.64462179 179 nips-2010-Natural Policy Gradient Methods with Parameter-based Exploration for Control Tasks

8 0.62249935 130 nips-2010-Interval Estimation for Reinforcement-Learning Algorithms in Continuous-State Domains

9 0.61221135 152 nips-2010-Learning from Logged Implicit Exploration Data

10 0.6091162 189 nips-2010-On a Connection between Importance Sampling and the Likelihood Ratio Policy Gradient

11 0.60821128 160 nips-2010-Linear Complementarity for Regularized Policy Evaluation and Improvement

12 0.59996051 4 nips-2010-A Computational Decision Theory for Interactive Assistants

13 0.56005341 208 nips-2010-Policy gradients in linearly-solvable MDPs

14 0.51846558 11 nips-2010-A POMDP Extension with Belief-dependent Rewards

15 0.50597048 78 nips-2010-Error Propagation for Approximate Policy and Value Iteration

16 0.49936929 168 nips-2010-Monte-Carlo Planning in Large POMDPs

17 0.49857652 212 nips-2010-Predictive State Temporal Difference Learning

18 0.49099579 37 nips-2010-Basis Construction from Power Series Expansions of Value Functions

19 0.48691112 196 nips-2010-Online Markov Decision Processes under Bandit Feedback

20 0.47936499 201 nips-2010-PAC-Bayesian Model Selection for Reinforcement Learning

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(0, 0.011), (13, 0.025), (27, 0.053), (29, 0.031), (30, 0.04), (39, 0.011), (45, 0.235), (50, 0.042), (52, 0.017), (60, 0.034), (77, 0.047), (78, 0.017), (85, 0.321), (90, 0.023)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.79044276 43 nips-2010-Bootstrapping Apprenticeship Learning

Author: Abdeslam Boularias, Brahim Chaib-draa

2 0.71372527 117 nips-2010-Identifying graph-structured activation patterns in networks

Author: James Sharpnack, Aarti Singh

Abstract: We consider the problem of identifying an activation pattern in a complex, largescale network that is embedded in very noisy measurements. This problem is relevant to several applications, such as identifying traces of a biochemical spread by a sensor network, expression levels of genes, and anomalous activity or congestion in the Internet. Extracting such patterns is a challenging task specially if the network is large (pattern is very high-dimensional) and the noise is so excessive that it masks the activity at any single node. However, typically there are statistical dependencies in the network activation process that can be leveraged to fuse the measurements of multiple nodes and enable reliable extraction of highdimensional noisy patterns. In this paper, we analyze an estimator based on the graph Laplacian eigenbasis, and establish the limits of mean square error recovery of noisy patterns arising from a probabilistic (Gaussian or Ising) model based on an arbitrary graph structure. We consider both deterministic and probabilistic network evolution models, and our results indicate that by leveraging the network interaction structure, it is possible to consistently recover high-dimensional patterns even when the noise variance increases with network size. 1

3 0.64050007 93 nips-2010-Feature Construction for Inverse Reinforcement Learning

Author: Sergey Levine, Zoran Popovic, Vladlen Koltun

4 0.61542159 63 nips-2010-Distributed Dual Averaging In Networks

Author: Alekh Agarwal, Martin J. Wainwright, John C. Duchi

Abstract: The goal of decentralized optimization over a network is to optimize a global objective formed by a sum of local (possibly nonsmooth) convex functions using only local computation and communication. We develop and analyze distributed algorithms based on dual averaging of subgradients, and provide sharp bounds on their convergence rates as a function of the network size and topology. Our analysis clearly separates the convergence of the optimization algorithm itself from the effects of communication constraints arising from the network structure. We show that the number of iterations required by our algorithm scales inversely in the spectral gap of the network. The sharpness of this prediction is conﬁrmed both by theoretical lower bounds and simulations for various networks. 1

5 0.61465347 239 nips-2010-Sidestepping Intractable Inference with Structured Ensemble Cascades

Author: David Weiss, Benjamin Sapp, Ben Taskar

Abstract: For many structured prediction problems, complex models often require adopting approximate inference techniques such as variational methods or sampling, which generally provide no satisfactory accuracy guarantees. In this work, we propose sidestepping intractable inference altogether by learning ensembles of tractable sub-models as part of a structured prediction cascade. We focus in particular on problems with high-treewidth and large state-spaces, which occur in many computer vision tasks. Unlike other variational methods, our ensembles do not enforce agreement between sub-models, but ﬁlter the space of possible outputs by simply adding and thresholding the max-marginals of each constituent model. Our framework jointly estimates parameters for all models in the ensemble for each level of the cascade by minimizing a novel, convex loss function, yet requires only a linear increase in computation over learning or inference in a single tractable sub-model. We provide a generalization bound on the ﬁltering loss of the ensemble as a theoretical justiﬁcation of our approach, and we evaluate our method on both synthetic data and the task of estimating articulated human pose from challenging videos. We ﬁnd that our approach signiﬁcantly outperforms loopy belief propagation on the synthetic data and a state-of-the-art model on the pose estimation/tracking problem. 1

6 0.614586 277 nips-2010-Two-Layer Generalization Analysis for Ranking Using Rademacher Average

7 0.6139822 1 nips-2010-(RF)^2 -- Random Forest Random Field

8 0.61384541 164 nips-2010-MAP Estimation for Graphical Models by Likelihood Maximization

9 0.61361814 282 nips-2010-Variable margin losses for classifier design

10 0.61317337 158 nips-2010-Learning via Gaussian Herding

11 0.61316431 287 nips-2010-Worst-Case Linear Discriminant Analysis

12 0.61294454 169 nips-2010-More data means less inference: A pseudo-max approach to structured learning

13 0.61293948 29 nips-2010-An Approximate Inference Approach to Temporal Optimization in Optimal Control

14 0.61115038 211 nips-2010-Predicting Execution Time of Computer Programs Using Sparse Polynomial Regression

15 0.61110741 177 nips-2010-Multitask Learning without Label Correspondences

16 0.61109424 224 nips-2010-Regularized estimation of image statistics by Score Matching

17 0.61094558 218 nips-2010-Probabilistic latent variable models for distinguishing between cause and effect

18 0.61070538 49 nips-2010-Computing Marginal Distributions over Continuous Markov Networks for Statistical Relational Learning

19 0.61066347 52 nips-2010-Convex Multiple-Instance Learning by Estimating Likelihood Ratio

20 0.61029416 236 nips-2010-Semi-Supervised Learning with Adversarially Missing Label Information