Sequential Decision Making

Transcription

1 Sequential Decision Making Dynamic programming Christos Dimitrakakis Intelligent Autonomous Systems, IvI, University of Amsterdam, The Netherlands March 18, 2008

2 Introduction Some examples Dynamic programming Summary

3 The purpose of this lecture Basic concepts Refresh memory. Present the MDP setting. Define optimality. Categorize planning tasks Algorithms Introduce basic planning algorithms. Promote intuition about their relationships. Discuss their applicability. Ultimate goal A firm foundation in reasoning and planning under uncertainty.

4 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

5 Preliminaries Variables Environment µ M States s t S. Actions a t A. A reward r t R. A policy π P. Notation Probabilities P(x y, z) z(x y). Expectations E(x y, z) Sometimes P(a t = a ) will be used for clarity. i.e. π t(a s) = P(a t = a s t = s, π t)

6 Markov decision processes The setting We are in some dynamic environment µ, where at each time step t we observe States s t S. Actions a t A. µ r t+1 s t s t+1 A reward r t R. a t P(s t+1 s t, a t, s t 1, a t 1,..., µ) = P(s t+1 s t, a t, µ) (1) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t+1, s t, a t, µ) (2)

7 Markov decision processes The setting We are in some dynamic environment µ, where at each time step t we observe States s t S. Actions a t A. µ r t+1 s t s t+1 A reward r t R. a t P(s t+1 s t, a t, s t 1, a t 1,..., µ) = P(s t+1 s t, a t, µ) (1) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t+1, s t, a t, µ) (2) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t, a t, µ) (3)

8 Markov decision processes The setting We are in some dynamic environment µ, where at each time step t we observe States s t S. Actions a t A. µ r t+1 s t s t+1 A reward r t R. a t P(s t+1 s t, a t, s t 1, a t 1,..., µ) = P(s t+1 s t, a t, µ) (1) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t+1, s t, a t, µ) (2) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t, a t, µ) (3) p(r t+1 s t+1, s t, a t, s t 1, a t 1,..., µ) = p(r t+1 s t+1, µ) (4)

9 Markov decision processes Controlling the environment We wish to control the environment according to some (for now undefined) optimality criterion. The agent The agent is fully defined by its policy π. This induces a probability distribution on actions and states. µ s t s t+1 a t π r t+1 P(a t s t, a t 2, s t 1, a t 2,..., π, µ) = P(a t s t, π) (5)

10 Markov decision processes µ The induced Markov chain Together with the policy π and the model µ, we induce a Markov chain on states. r t+1 s t s t+1 a t P(s t+1 s t, π, µ) = X a A P(s t+1 a t = a, s t, π, µ) P(a t = a s t, π) (6a) π P(s t+k s t, π, µ) = X s P(s t+k s t+k 1 = s, π, µ) P(s t+k 1 s t, π, µ) (6b) Note: lim k P(s t+k = s s t, π, µ) is the stationary distribution.

11 Markov decision processes µ The induced Markov chain Together with the policy π and the model µ, we induce a Markov chain on states. r t+1 s t s t+1 P(s t+1 s t, π, µ) = X a A P(s t+1 a t = a, s t, π, µ) P(a t = a s t, π) (6a) π P(s t+k s t, π, µ) = X s P(s t+k s t+k 1 = s, π, µ) P(s t+k 1 s t, π, µ) (6b) Note: lim k P(s t+k = s s t, π, µ) is the stationary distribution.

12 Planning The goal in reinforcement learning To maximise a function of future rewards. Finite horizon We are only interested in rewards up to a fixed point in time. Infinite horizon We are interested in all rewards.

13 Value functions The return / utility The agent s goal is to maximize the return (Too many Rs, switching to U). For example the utility given a policy π and an MDP µ! TX Ut,µ( π ) E(U, π, µ) = E γ k r t+k, π, µ (7) TX = γ X k E[r t+k s t+k =i, µ] P(s t+k = i, π, µ) (8) k=1 i S Can in principle be calculated from (6). The value functions k=1 Special case: T, V π t (s) = V π (s). V π t (s) X a A U π t,µ(s, a)π(a s) (9) Q π t (s, a) U π t,µ(s, a) (10)

14 Bellman equation An optimal policy An optimal policy has the property that whatever the initial state and initial decision are, the remaining decisions must constitute an optimal policy with regard to the state resulting from the first decision. The recursion TX t t (s) = g(t) E[r t+1 s t=s, π] + g(t + k) E[r t+k s t=s, a t=a, π, µ] (11) V π k=2 = g(t) E[r t+1 s t=s, π] + X i S V π t+1(i)µ(s t+1=i s t=s, π). (12) The current stage s value is just the next reward plus the next stage s value. See also the Hamilton-Jacobi-Bellman equation in optimal control.

15 Greedy policies The 1-step greedy policy The 1-step-greedy policy with respect to a given value function can be expressed as ( 1, a = arg max π(a s) = a Q(s, a ) 0, otherwise (13) The optimal policy The 1-step-greedy policy with respect to the optimal value function is optimal. Naive solution Evaluate all policies, select π : V π (s) V π (s) s S. Clever solutions Directly estimate V. Iteratively improve π.

16 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

17 Problem types Planning with... Finite vs Infinite horizon Discounted vs Undiscounted rewards Certain vs Uncertain knowledge Expected vs worst-case utility functions Environments Deterministic Stochastic Episodic Continuing Observable Hidden state Statistical Adversarial

18 Deterministic shortest-path problems X Properties g(t) = 1, T. r t = 1 unless s t = X, in which case r t = 0. µ(s t+1 = X s t = X ) = 1. A = {North, South, East, West} Transitions are deterministic and walls block. What is the shortest path to the destination from any point?

19 Stochastic shortest path problem, with a pit O X Properties g(t) = 1, T. r t = 1, but r t = 0 at X and 100 at O and episode ends. µ(s t+1 = X s t = X ) = 1. A = {North, South, East, West} Moves to a random direction with probability θ. Walls block. For what value of θ is it better to take the dangerous shortcut? (However, if we want to take into account risk explicitly we must modify the agent s utility function)

20 Continuing stochastic MDPs Inventory management There are K storage locations. Each place can store n i items. At each time-step there is a probability φ i that a client try to buy an item from location i, P i φ i 1. If there is an item available, you gain reward 1. Action 1: ordering u units of stock, for paying c(u). Action 2: move u units of stock from one location i to another, j, for a cost ψ ij (u). An easy special case K = 1. There is one type of item only. Orders are placed and received every n timesteps.

21 Inventory management An easy special case K = 1. Deliveries happen once every m timesteps. Each time-step a client arrives with probability φ. Properties The state set. The action set. The transition probabilities

22 Inventory management An easy special case K = 1. Deliveries happen once every m timesteps. Each time-step a client arrives with probability φ. Properties The state set is the number of items we have: S = {0, 1,..., n}. The action set. The transition probabilities

23 Inventory management An easy special case K = 1. Deliveries happen once every m timesteps. Each time-step a client arrives with probability φ. Properties The state set is the number of items we have: S = {0, 1,..., n}. The action set A = {0, 1,..., n} since we can order from nothing up to n items. The transition probabilities

24 Inventory management An easy special case K = 1. Deliveries happen once every m timesteps. Each time-step a client arrives with probability φ. Properties The state set is the number of items we have: S = {0, 1,..., n}. The action set A = {0, 1,..., n} since we can order from nothing up to n items. The transition probabilities P(s s, a) = `m d φd (1 φ) m d, where d = s + a s, for s + a n.

25 Episodic, finite, infinite? Shortest path problems Episodic tasks with infinite horizon, 1 reward everywhere, but 0 in absorbing state. Continuing tasks with 0 reward everywhere, but > 0 in goal state, γ (0, 1), state reset after goal. Equivalent if optimal policy is the same.

26 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

27 Introduction Why dynamic programming? Programming means finding a solution. i.e. linear programming. Dynamic because we find solution to dynamical problems. Direct relation to control theory.

28 The shortest-path problem revisited Properties γ = 1, T. r t = 1 unless s t = X, in which case r t = 0. The length of the shortest path from s equals the negative value of the optimal policy. Also called cost-to-go. Remember Dijkstra s algorithm?

29 Backwards induction I s 4 T 2 st 3 2 st 2 2 s 2 T 1 s 1 T 1 s T If we know the value of the last state, we can calculate the values of its predecessors. The value of s i T 1 is the reward obtained by moving from s i T 1 to s T, plus the value of s T. s 1 T 2

30 Backwards induction II w B y D max{w + y, z + x + w} A B C D 0 w w e x y A B 0 w A 0 w A x z C x + w All w, x, y, z < 0, and reward e < 0 of staying at the same state, apart from A. All w, x, y, z

31 Backwards induction III Backwards induction in deterministic environments Input µ, S T. Initialise V T (s), for all s S T. for n = T 1, T 2,..., t do for s S n do a n (s) = arg max a E(r s s,a, s, µ) + V n+1(s s,a) V n (s) = E(r s s,a n (s), s, µ) + V n+1(s s,a n (s) ) end for end for Notes s s,a is the state that occurs if we take a in s. Because we always know the optimal choice at the last step, we can find the optimal policy directly!

32 Backwards induction III Backwards induction in deterministic environments Input µ, S T. Initialise V T (s), for all s S T. for n = T 1, T 2,..., t do for s S n do a n (s) = arg max a Ps S n+1 µ(s s, a) E(r s, s, µ) + V n+1(s ) V n(s) = P s S n+1 µ(s s, a n (s)) E(r s, s, µ) + V n+1(s ) end for end for Notes µ(s s, a) is an indicator function Because we always know the optimal choice at the last step, we can find the optimal policy directly!

33 Backwards induction III Backwards induction in deterministic environments Input µ, S T. Initialise V T (s), for all s S T. for n = T 1, T 2,..., t do for s S n do a n (s) = arg max a Ps S n+1 µ(s s, a) E(r s, s, µ) + V n+1(s ) V n(s) = P s S n+1 µ(s s, a n (s)) E(r s, s, µ) + V n+1(s ) end for end for Notes µ(s s, a) is an indicator function Nothing apparently stops µ(s s, a) from being a distribution So, what happens in stochastic environments?

34 Backwards induction IV: Stochastic problems a 0 0 A a 1 0 A B a 0 0 w w B Almost as before, but state depends stochastically on actions, i.e. µ(s t+1=a s t=b, a t=a) a 1 The backup operators V π n (s) = X s [µ(s s, π) E(r s, s) + V π n+1(s )] (14) V n (s) = max a X s µ(s s, a)[e(r s, s) + V n+1(s )] (15)

35 Backwards induction V Policy evaluation with Backwards induction Input π, µ, S T. Initialise V T (s), for all s S T. for n = T 1, T 2,..., t do for s S n do V π n (s) = P s S n+1 µ(s s, π)[e(r s, s, µ) + V π n+1(s )] end for end for Notes µ(s s, π) = P a µ(s s, a)π(a s). Finite horizon problems only, or approximations to finite horizon (i.e. lookahead in game trees). Hey, it works for stochastic problems too! (By marginalizing over states) Because we always know the optimal choice at the last step, we can find the optimal policy directly! Can be used with estimates of the value function.

36 Backwards induction V Finding the optimal policy with Backwards induction Input µ, S T. Initialise V T (s), for all s S T. for n = T 1, T 2,..., t do for s S n do a n (s) = arg max a µ(s s, a)[e(r s, s, µ) + V n+1(s )] V n(s) = P s S n+1 µ(s s, a n )[E(r s, s, µ) + V n+1(s )] end for end for Notes Finite horizon problems only, or approximations to finite horizon (i.e. lookahead in game trees). Hey, it works for stochastic problems too! (By marginalizing over states) Because we always know the optimal choice at the last step, we can find the optimal policy directly! Can be used with estimates of the value function.

37 Infinite horizon What happens when the horizon is infinite in stochastic shortest path problems? Episodic tasks still terminate with probability one for proper policies. Assumption: there exists at least one proper policy. Assumption: Every improper policy has negatively infinite value for at least one state.

38 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

39 Policy improvement Why evaluate a policy? We can always generate a better policy given the value function of any policy! Theorem (Policy improvement) Let some policy π P. If π (a s) = 1 for a = arg max a Q π (s, a) and 0 otherwise, then V π (s) V π (s), s S

40 Policy improvement theorem Theorem (Policy improvement) Let some policy π P. If π (a s) = 1 for a = arg max a Q π (s, a) and 0 otherwise, then V π (s) V π (s), s S Proof. Let π k be the policy which execute π for k steps and then reverts to π. Then π = π 0, π = lim k π k, and we have V π (s t) = X a t π(a t s t)q π (s, a) 2 3 max a t Q π (s, a) = max a t 4 X µ(s t+1 s t, a t)v π (s t+1) 5 = V π 1 (s t). st+1 Similarly, we show that V π k+1 (s) V π k (s) for all s. Then V π V π 1 (s) V π k (s) V π k+1 (s)... and so V π (s) = lim k V π k (s) V π (s).

41 Iterative policy evaluation Policy Evaluation Input π, µ and ˆV 0. n = 0. repeat n = n + 1 for s S do ˆV n(s) = P a A π(a s) P s S µ(s s, a)[e(r s, µ) + γ ˆV n 1(s )] end for until ˆV n ˆV n 1 < θ Notes Arbitrary initialization. V π, ˆV n R S, lim n ˆV n = V π, if the limit exists. Can be done in-place as well.

42 Policy evaluation example I iterations Random policy evaluation.

43 Policy evaluation example I iteration Random policy evaluation.

44 Policy evaluation example I iterations Random policy evaluation.

45 Policy evaluation example I iterations Random policy evaluation.

46 Policy evaluation example I Greedy policy with respect to value function of random policy Random policy evaluation.

47 Policy evaluation example II Random policy evaluation.

48 Policy evaluation example II Random policy evaluation.

49 Policy evaluation example II Random policy evaluation.

50 Policy evaluation example II Random policy evaluation.

51 Value iteration Value Iteration Input µ. ˆV 0(s) = 0 for all s S. n = 0. repeat n = n + 1 for s S do ˆV n(s) = max a A Ps S µ(s s, a)[e(r s, µ) + γ ˆV n 1(s )] end for until ˆV n ˆV n 1 < θ Notes No reason to assume a fixed policy, convergence holds. lim n ˆVn = V. Equivalent to backwards induction as horizon. This is because lim T V π t (s) = V π (s) for all t.

52 Value iteration example iter: 0

53 Value iteration example iter: 1

54 Value iteration example iter: 10

55 Value iteration example iter: 100

56 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

57 Policy iteration I Policy Iteration Input π, µ. repeat Evaluate V π. π : π (s) = arg max a Q π (s, a) until arg max a Q π (s, a) = V π (s) for all s Theorem (Policy iteration) The policy iteration algorithm generates an improving sequence of proper policies, i.e. V π k+1 (s) V π k (s), k > 0, s S and terminates with an optimal policy, i.e. lim k V π k = V. Remark (Policy iteration termination) If π k is not optimal, then s S : V π k+1 (s) > V π k (s). Conversely, if no such s exists, π k is optimal and we terminate.

58 Policy iteration II The evaluation step It can be done exactly by solving the linear equations. (Proper policy iteration) We can use a limited number n of policy evaluation iterations (Modified policy iteration algorithm). These can be initalised from the last evaluation. If we use just n = 1, then the method is identical to value iteration. If we use n, then we have proper policy iteration. Other methods Asynchronous policy iteration. Multistage lookahead policy iteration. See [1], section 2.2 for more details. See [3], Chapters 4,5,6 for detailed theory.

59 Preliminaries Markoc decision processes Value functions and optimality Introduction Some examples Shortest-path problems Continuing problems Episodic, finite, infinite? Dynamic programming Introduction Backwards induction Iterative Methods Policy evaluation Value iteration Policy iteration Summary Lessons learnt Learning from reinforcement... Bibliography

60 Lessons learnt Planning with a known model Find the optimal policy given model and objective. Bellman recursion is the basis of dynamic programming. Easy to solve for finite-horizon problems or episodic tasks. Stochasticity does not make the problem significantly harder. Infinite-horizon continuing problems harder, but tractable. Things to think about Would iterative methods be better than backwards induction? How does it depend on the problem? Does the discount factor have any effect? How can backwards induction be applied to iterative problems and vice-versa?

61 Learning from reinforcement... Bandit problems γ [0, 1], T > 0. S = 1. Rewards are random with expectation E[r t a t, µ] If µ known, trivial: a = arg max a E[r t a t = a, µ], for all t, γ. If µ is unknown, can be intractable. Simplest case of learning from reinforcement.

62 Further reading Dimitri P. Bertsekas and John N. Tsitsiklis. Neuro-Dynamic Programming. Athena Scientific, Morris H. DeGroot. Optimal Statistical Decisions. John Wiley & Sons, Republished in Marting L. Puterman. Markov Decision Processes : Discrete Stochastic Dynamic Programming. John Wiley & Sons, New Jersey, US, 1994,2005. Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. MIT Press, 1998.