Reinforcement Learning

Transcription

1 Reinforcement Learning Model-Based Reinforcement Learning Model-based, PAC-MDP, sample complexity, exploration/exploitation, RMAX, E3, Bayes-optimal, Bayesian RL, model learning Vien Ngo MLR, University of Stuttgart

2 Outline Model-Based RL Exploration/Exploitation PAC-MDP E 3, RMAX Bayes optimal: Bayesian model-based RL 2/??

3 RL Approaches 3/??

4 Model-based RL Model learning: Given data D = {(s t, a t, r t, s t+1 )} H t=1 estimate P (s a, s) and P (r a, s) discrete state-action: ˆP (s a, s) = #(s,a,s) #(a,s) continuous state-action: ˆP (s a, s) = N (s φ(s, a) β, Σ) estimate parameters β (and perhaps Σ) as for regression (including non-linear features, regularization, cross-validation!) Planning: discrete state-action: Dynamic Programming with the estimated model continuous state-action: Differential Dynamic Programming, Planning-by-Inference, etc. 4/??

5 Exploration vs. Exploitation 5/??

6 Exploration: Motivation Start: many potential paths 6/??

7 Exploration: Motivation Continue using this path? Try out alternatives? 7/??

8 Exploration: Motivation 8/??

9 Exploration: Motivation 9/??

10 Reinforcement learning In reinforcement learning (RL), the agent starts to act without a model of the environment. The agent has to learn from its experience what to do to in order to fulfill tasks and achieve high rewards. RL algorithms we have seen thus far: Q-learning, TD-learning, policy gradient,... 10/??

11 Reinforcement learning In reinforcement learning (RL), the agent starts to act without a model of the environment. The agent has to learn from its experience what to do to in order to fulfill tasks and achieve high rewards. RL algorithms we have seen thus far: Q-learning, TD-learning, policy gradient,... Note the difference to the problem of adapting the behavior based on a given model (also called planning / solving an MDP / calculating optimal state and action values). This is a computational subproblem in model-based RL. Planning algorithms we have seen thus far: value iteration, policy iteration. 10/??

12 Exploration / Exploitation In contrast to supervised learning, in RL the data used for learning depend on the agent. Two different types of behavior: exploration: behave with the goal to learn as much as possible exploitation: behave with the goal of getting as much reward as possible Challenge in exploration: which actions will lead to the most important learning progress with respect to the goal? exploration as fundamental intelligent behavior 11/??

13 Recall Markov Decision Processes a 0 a 1 a 2 s 0 s 1 s 2 r 0 r 1 r 2 P (s 0:T, a 0:T, r 0:T ; π) = P (s 0 )P (a 0 s 0 ; π)p (r 0 a 0, s 0 ) T t=1 P (st a t-1, s t-1 )P (a t s t; π)p (r t a t, s t) world s initial state distribution P (s 0 ) world s transition probabilities P (s t+1 a t, s t ) world s reward probabilities P (r t a t, s t ) agent s policy π(a t s t ) (or deterministic a t = π(s t )) discount parameter γ for future rewards two different sources of uncertainty: the world itself (not controlled by the agent) vs. the policy (controlled by the agent) 12/??

14 Exploration-exploitation tradeoff Goal of reinforcement learning agent: maximize future rewards E[ t=0 γt r t s 0 ; π] However, the agent does not know the transition parameters P (s t+1 a t, s t ) and reward parameters P (r t a t, s t ) of the MDP. Rather, the agent needs to learn from its experience s 0, a 0, r 0, s 1, a 1, r 1,... which actions will lead to high rewards. 13/??

15 Exploration-exploitation tradeoff Goal of reinforcement learning agent: maximize future rewards E[ t=0 γt r t s 0 ; π] However, the agent does not know the transition parameters P (s t+1 a t, s t ) and reward parameters P (r t a t, s t ) of the MDP. Rather, the agent needs to learn from its experience s 0, a 0, r 0, s 1, a 1, r 1,... which actions will lead to high rewards. Exploration-exploitation tradeoff: Which policy π(a t s t ) for action selection shall the agent follow so that it does not miss the high-reward states, but does not spend too much time in low-reward states, either? Exploitation: Prefer actions a t which have led to reward before? Exploration: Or rather take actions to learn more about the unknown MDP parameters and potentially find states with higher reward? 13/??

16 Notion of Optimality 14/??

17 Sample Complexity Let M be an MDP with N states, K actions, discount factor γ [0, 1) and a maximal reward R max > 0. Let A be an algorithm (that is, a reinforcement learning agent) that acts in the environment, resulting in s 0, a 0, r 0, s 1, a 1, r 1,.... Let V A t,m = E[ i=0 γi r t+i s 0, a 0, r 0... s t 1, a t 1, r t 1, s t ]. V is the value function of the optimal policy. Define an accuracy threshold ɛ: ˆV V ɛ 15/??

18 Definition: Let ɛ > 0 be a prescribed accuracy and δ > 0 be an allowed probability of failure. The expression η(ɛ, δ, N, K, γ, R max ) is a sample complexity bound for algorithm A if independently of the choice of s 0, with probability at least 1 δ, the number of timesteps such that V A t,m < V (s t ) ɛ is at most η(ɛ, δ, N, K, γ, R max ). (Kakade, 2003) 16/??

19 Efficient exploration An algorithm with sample complexity that is polynomial in 1/ɛ, log(1/δ), N, K, 1/(1 γ), R max is called PAC-MDP (probably approximately correct in MDPs). 17/??

20 Exploration strategies 18/??

21 Exploration strategies The exploration strategy is reflected in the policy π(s). In the following, assume we have estimates ˆQ(s, a). greedy (only exploit): π(s) = argmax a ˆQ(s, a) problem: learned model not the same as the true environment Without exploration, agent is likely to miss high rewards. random: choose action a with probability 1/{ actions} problem: ignores value estimates and thus rewards 19/??

22 Exploration strategies (continued) argmax a ˆQ(s, a) with probability 1 ɛ ɛ-greedy: π(s) = random action with probability ɛ most popular method Converges to the optimal value function with probability 1 (all paths will be visited sooner or later), if the exploration rate diminishes according to an appropriate schedule. problem: sample complexity exponential in the number of states exp( Boltzmann: choose action a with probability ˆQ(s,a)/T ) a exp( ˆQ(s,a)/T ) temperature T controls amount of exploration problem again: sample complexity exponential in the number of states 20/??

23 Exploration strategies (continued) Other heuristics for exploration: minimize variance of action value estimates optimistic initial values ( optimism in the face of uncertainty ) state bonuses: frequency, recency, error etc. problem again: sample complexity exponential in the number of states Bayesian RL: optimal exploration strategy distribution over MDP models (i.e., the parameters of the MDP) posterior distribution updated after each new observation exploration strategy minimizes uncertainty of parameters Bayes-optimal solution to the exploration-exploitation tradeoff (i.e., no other policy is better in expectation w.r.t. prior distribution over MDPs) only tractable for very simple problems E 3 and R max : principled approach to the exploration-exploitation tradeoff with polynomial sample complexity 21/??

24 PAC-MDP Algorithms Explicit-Exploit-or-Explore (E3) RMAX 22/??

25 PAC-MDP Algorithms Explicit-Exploit-or-Explore (E3) RMAX Common Instuition: optimism in the face of uncertainty (i.e. if faced the option of certain and uncertain reward regions, try the UNCERTAIN reward region) 22/??

26 Explicit-Exploit-or-Explore (E3) algorithm 23/??

27 Explicit-Exploit-or-Explore (E3) algorithm Kearns and Singh (2002) Model-based approach with polynomial sample complexity (PAC-MDP) uses optimism in the face of uncertainty assumes knowledge of maximum reward Maintains counts for states and actions to quantify confidence in model estimates A state s is known if all actions in s have been sufficiently often executed. 24/??

28 Explicit-Exploit-or-Explore (E3) algorithm Kearns and Singh (2002) Model-based approach with polynomial sample complexity (PAC-MDP) uses optimism in the face of uncertainty assumes knowledge of maximum reward Maintains counts for states and actions to quantify confidence in model estimates A state s is known if all actions in s have been sufficiently often executed. From observed data, E3 constructs two MDPs: MDP known : includes known states with (approximately exact) estimates of P (s t+1 a t, s t ) and P (r t a t, s t ) model which captures what you know MDP unknown : MDP known + special state s where the agent receives maximum reward model which drives exploration 24/??

29 E3 sketch Input: State s Output: Action a if s is known then Plan in MDP known Sufficiently accurate model estimates if resulting plan has value above some threshold then return first action of plan Exploitation else Plan in MDP unknown return first action of plan Planned exploration else return action with the least observations in s Direct exploration 25/??

30 E3 example S. Singh (Tutorial 2005) 26/??

31 E3 example S. Singh (Tutorial 2005) 27/??

32 E3 example M : true known state MDP S. Singh (Tutorial 2005) M : estimated known state MDP 28/??

33 E 3 T is the time horizon. Implementation Setting G T max is the maximum T-step return. (discounted case G T max T R max ). ( ) A state is known if it was visited O (NT G T max/ɛ) 4 V ar max log(1/δ) times. (V ar max is the maximum variance of the random payoffs over all states). For the exploration/exploitation choice at known states: It s assumed to be given the optimal value function V. If ˆV obtained from the MDP known > (V ɛ) then do exploitation. 29/??

34 RMAX Algorithm 30/??

35 RMAX R-MAX solves only one unique model (don t separate MDP known and MDP unknown ) and therefore implicitly explores or exploits. The R-MAX and E3 algorithms were able to achieve roughly the same level of performance (Strehl s thesis). 31/??

36 RMAX R-MAX solves only one unique model (don t separate MDP known and MDP unknown ) and therefore implicitly explores or exploits. The R-MAX and E3 algorithms were able to achieve roughly the same level of performance (Strehl s thesis). RMAX builds an approximate MDP based on reward function ˆR(s, a) R(s, a) = R max if (s,a) known otherwise 31/??

37 RMAX s sketch Initialize all couter n(s, a) = 0, n(s, a, s ) = 0. Initialize ˆT (s s, a) = I s=s, ˆR(s, a) = R max while (1) do Compute policy π t using MDP model of ( ˆT, ˆR). Choose a = π t(s), observe s, r. n(s, a) = n(s, a) + 1 r(s, a) = r(s, a) + r n(s, a, s ) = n(s, a, s ) + 1 if n(s, a) = m then Update ˆT ( s, a) = n(s, a, )/n(s, a), and ˆR(s, a) = r(s, a)/n(s, a). 32/??

38 RMAX Analysis The general PAC-MDP theorem does not easily adapt to the analysis of E3 because of E3 s use of two internal models (Original analysis depends on horizon and mixing time). ( (Upper bound) There exists m = O NT 2 ɛ 2 of at least 1 δ, V (s t ) V (s t ) ɛ is true for all but ( N 2 KT 3 O ɛ 3 where N is the number of states. For discounted case: T = log ɛ 1 γ ) ln 2 NK δ, then with probability ) ln 2 NK δ 33/??

39 Limitations of E3 and RMAX E3/RMAX is called efficient because its sample complexity scales only polynomially in the number of states. In natural environments, however, this number of states is enormous: it is exponential in the number of objects in the environment. Hence E3/RMAX scales exponentially in the number of objects. Generalization over states and actions is crucial for exploration. 34/??

40 KWIK KWIK (Known What It Knows): a supervised-learning model 35/??

41 KWIK KWIK (Known What It Knows): a supervised-learning model Input set: X Output set: Y Observation set: Z Hypothesis class: H : X Y Target function: h H Special symbol:? (I don t know) 36/??

42 KWIK 37/??

43 KWIK Example: Coin Learning Predict P r(head) [0, 1] for a coin from many observations: head or tail Algorithm Predict? the first O(1/ɛ 2 log(1/δ)) times Use empirical estimate afterwards The bound follows from Hoeffdings bound O(1/ɛ 2 log(1/δ)) Li et. al. ICML /??

44 KWIK-RMAX T (s s, a) and R(s, a) are TWIK-learned. 39/??

45 Conclusions RL agents need to solve the exploration-exploitation tradeoff. Sample complexity measures the required number of explorative actions of an algorithm. Ideas for driving exploration: random actions, optimism in the face of uncertainty, maximizing learning progress and information gain 40/??

46 References Kakade (2003): On the sample complexity of reinforcement learning. PhD thesis. Poupart, Vlassis, Hoey, Kevin Regan: An analytic solution to discrete Bayesian reinforcement learning. ICML Li, Littman, Walsh: Knows what it knows: a framework for self-aware learning. ICML /??