Presentation is loading. Please wait.

Presentation is loading. Please wait.

CS 541: Artificial Intelligence Lecture X: Markov Decision Process Slides Credit: Peter Norvig and Sebastian Thrun.

Similar presentations


Presentation on theme: "CS 541: Artificial Intelligence Lecture X: Markov Decision Process Slides Credit: Peter Norvig and Sebastian Thrun."— Presentation transcript:

1

2 CS 541: Artificial Intelligence Lecture X: Markov Decision Process Slides Credit: Peter Norvig and Sebastian Thrun

3 Schedule  TA: Vishakha Sharma  Email Address: vsharma1@stevens.edu WeekDateTopicReadingHomework 108/30/2011Introduction & Intelligent AgentCh 1 & 2N/A 209/06/2011Search: search strategy and heuristic searchCh 3 & 4sHW1 (Search) 309/13/2011Search: Constraint Satisfaction & Adversarial SearchCh 4s & 5 & 6 Teaming Due 409/20/2011Logic: Logic Agent & First Order LogicCh 7 & 8sHW1 due, Midterm Project (Game) 509/27/2011Logic: Inference on First Order LogicCh 8s & 9 610/04/2011Uncertainty and Bayesian NetworkCh 13 &14sHW2 (Logic) 10/11/2011No class (function as Monday) 710/18/2011Midterm Presentation Midterm Project Due 810/25/2011Inference in Baysian NetworkCh 14sHW2 Due, HW3 (Probabilistic Reasoning) 911/01/2011Probabilistic Reasoning over TimeCh 15 1011/08/2011TA Session: Review of HW1 & 2 HW3 due, HW4 (MDP) 1111/15/2011Supervised Machine LearningCh 18 1211/22/2011Markov Decision ProcessCh 16 HW4 due 1311/29/2011Reinforcement learningCh 21 1412/06/2011Final Project Presentation Final Project Due

4 Re-cap of Lecture IX  Machine Learning  Classification (Naïve Bayes)  Regression (Linear, Smoothing)  Linear Separation (Perceptron, SVMs)  Non-parametric classification (KNN)

5 Outline  Planning under uncertainty  Planning with Markov Decision Processes  Utilities and Value Iteration  Partially Observable Markov Decision Processes

6 Planning under uncertainty

7 Rhino Museum Tourguide 6 http://www.youtube.com/watch?v=PF2qZ-O3yAM

8 Minerva Robot 7 http://www.cs.cmu.edu/~rll/resources/cmubg.mpg

9 Pearl Robot 8 http://www.cs.cmu.edu/~thrun/movies/pearl-assist.mpg

10 Mine Mapping Robot Groundhog 9 http://www-2.cs.cmu.edu/~thrun/3D/mines/groundhog/anim/mine_mapping.mpg

11 10

12 Planning Uncertainty 11

13 Planning: Classical Situation hellheaven World deterministic State observable

14 MDP-Style Planning hellheaven World stochastic State observable [Koditschek 87, Barto et al. 89] Policy Universal Plan Navigation function

15 Stochastic, Partially Observable sign hell?heaven? [Sondik 72] [Littman/Cassandra/Kaelbling 97]

16 Stochastic, Partially Observable sign hellheaven sign heavenhell

17 Stochastic, Partially Observable sign heavenhell sign ?? hellheaven start 50%

18 Stochastic, Partially Observable sign ?? start sign heavenhell sign hellheaven 50% sign ?? start

19 A Quiz  -dim continuous stochastic 1-dim continuous stochastic actions# statessize belief space?sensors 3: s 1, s 2, s 3 deterministic3perfect 3: s 1, s 2, s 3 stochastic3perfect 2 3 -1: s 1, s 2, s 3, s 12, s 13, s 23, s 123 deterministic3 abstract states deterministic3stochastic 2-dim continuous: p ( S=s 1 ), p ( S=s 2 ) stochastic3none 2-dim continuous: p ( S=s 1 ), p ( S=s 2 )  -dim continuous deterministic 1-dim continuous stochastic aargh! stochastic  -dim continuous stochastic

20 Planning with Markov Decision Processes

21 MPD Planning  Solution for Planning problem  Noisy controls  Perfect perception  Generates “universal plan” (=policy)

22 Grid World  The agent lives in a grid  Walls block the agent’s path  The agent’s actions do not always go as planned:  80% of the time, the action North takes the agent North (if there is no wall there)  10% of the time, North takes the agent West; 10% East  If there is a wall in the direction the agent would have been taken, the agent stays put  Small “living” reward each step  Big rewards come at the end  Goal: maximize sum of rewards*

23 Grid Futures 22 Deterministic Grid WorldStochastic Grid World X X E N S W X ? X XX

24 Markov Decision Processes  An MDP is defined by:  A set of states s  S  A set of actions a  A  A transition function T(s,a,s’)  Prob that a from s leads to s’  i.e., P(s’ | s,a)  Also called the model  A reward function R(s, a, s’)  Sometimes just R(s) or R(s’)  A start state (or distribution)  Maybe a terminal state  MDPs are a family of non-deterministic search problems  Reinforcement learning: MDPs where we don’t know the transition or reward functions 23

25 Solving MDPs  In deterministic single-agent search problems, want an optimal plan, or sequence of actions, from start to a goal  In an MDP, we want an optimal policy  *: S → A  A policy  gives an action for each state  An optimal policy maximizes expected utility if followed  Defines a reflex agent Optimal policy when R(s, a, s’) = -0.03 for all non- terminals s

26 Example Optimal Policies R(s) = -2.0 R(s) = -0.4 R(s) = -0.03R(s) = -0.01

27 MDP Search Trees  Each MDP state gives a search tree a s s’s’ s, a (s,a,s’) called a transition T(s,a,s’) = P(s’|s,a) R(s,a,s’) s,a,s’ s is a state (s, a) is a q- state 26

28 Why Not Search Trees?  Why not solve with conventional planning?  Problems:  This tree is usually infinite (why?)  Same states appear over and over (why?)  We would search once per state (why?) 27

29 Utilities

30  Utility = sum of future reward  Problem: infinite state sequences have infinite rewards  Solutions:  Finite horizon:  Terminate episodes after a fixed T steps (e.g. life)  Gives nonstationary policies (  depends on time left)  Absorbing state: guarantee that for every policy, a terminal state will eventually be reached (like “done” for High-Low)  Discounting: for 0 <  < 1  Smaller  means smaller “horizon” – shorter term focus 29

31 Discounting  Typically discount rewards by  < 1 each time step  Sooner rewards have higher utility than later rewards  Also helps the algorithms converge 30

32 Recap: Defining MDPs  Markov decision processes:  States S  Start state s 0  Actions A  Transitions P(s’|s,a) (or T(s,a,s’))  Rewards R(s,a,s’) (and discount  )  MDP quantities so far:  Policy = Choice of action for each state  Utility (or return) = sum of discounted rewards a s s, a s,a,s’ s’s’ 31

33 Optimal Utilities  Fundamental operation: compute the values (optimal expect-max utilities) of states s  Why? Optimal values define optimal policies!  Define the value of a state s: V * (s) = expected utility starting in s and acting optimally  Define the value of a q-state (s,a): Q * (s,a) = expected utility starting in s, taking action a and thereafter acting optimally  Define the optimal policy:  * (s) = optimal action from state s a s s, a s,a,s’ s’s’ 32

34 The Bellman Equations  Definition of “optimal utility” leads to a simple one-step lookahead relationship amongst optimal utility values: Optimal rewards = maximize over first action and then follow optimal policy  Formally: a s s, a s,a,s’ s’s’ 33

35 Solving MDPs  We want to find the optimal policy  *  Proposal 1: modified expect-max search, starting from each state s: a s s, a s,a,s’ s’s’ 34

36 Value Iteration  Idea:  Start with V 0 * (s) = 0  Given V i *, calculate the values for all states for depth i+1:  This is called a value update or Bellman update  Repeat until convergence  Theorem: will converge to unique optimal values  Basic idea: approximations get refined towards optimal values  Policy may converge long before values do 35

37 Example: Bellman Updates 36 max happens for a=right, other actions not shown Example:  =0.9, living reward=0, noise=0.2

38 Example: Value Iteration  Information propagates outward from terminal states and eventually all states have correct value estimates V2V2 V3V3 37

39 Computing Actions  Which action should we chose from state s:  Given optimal values V?  Given optimal q-values Q?  Lesson: actions are easier to select from Q’s!

40 Asynchronous Value Iteration*  In value iteration, we update every state in each iteration  Actually, any sequences of Bellman updates will converge if every state is visited infinitely often  In fact, we can update the policy as seldom or often as we like, and we will still converge  Idea: Update states whose value we expect to change: If is large then update predecessors of s

41 Value Iteration: Example

42 Another Example Value Function and PlanMap

43 Another Example Value Function and PlanMap

44 Partially Observable Markov Decision Processes

45 Stochastic, Partially Observable sign ?? start sign heavenhell sign hellheaven 50% sign ?? start

46 Value Iteration in Belief space: POMDPs  Partially Observable Markov Decision Process  Known model (learning even harder!)  Observation uncertainty  Usually also: transition uncertainty  Planning in belief space = space of all probability distributions  Value function: Piecewise linear, convex function over the belief space 45

47 Introduction to POMDPs (1 of 3) 80 100 ba 0 ba 40 s2s2 s1s1 action a action b p(s1)p(s1) [Sondik 72, Littman, Kaelbling, Cassandra ‘97] s2s2 s1s1 100 0 100 action aaction b

48 Introduction to POMDPs (2 of 3) 80 100 ba 0 ba 40 s2s2 s1s1 80% c 20% p(s1)p(s1) s2s2 s1’s1’ s1s1 s2’s2’ p(s1’)p(s1’) p(s1)p(s1) s2s2 s1s1 100 0 100 [Sondik 72, Littman, Kaelbling, Cassandra ‘97]

49 Introduction to POMDPs (3 of 3) 80 100 ba 0 ba 40 s2s2 s1s1 80% c 20% p(s1)p(s1) s2s2 s1s1 100 0 100 p(s1)p(s1) s2s2 s1s1 s1s1 s2s2 p(s 1 ’|A) B A 50% 30% 70% B A p(s 1 ’|B) [Sondik 72, Littman, Kaelbling, Cassandra ‘97]

50 POMDP Algorithm  Belief space = Space of all probability distribution (continuous)  Value function: Max of set of linear functions in belief space  Backup: Create new linear functions  Number of linear functions can grow fast! 49

51 Why is This So Complex? State Space Planning (no state uncertainty) Belief Space Planning (full state uncertainties) ?

52 but not usually. The controller may be globally uncertain... Belief Space Structure

53 Augmented MDPs: [Roy et al, 98/99] conventional state space uncertainty (entropy)

54 Path Planning with Augmented MDPs information gainConventional plannerProbabilistic Planner [Roy et al, 98/99]


Download ppt "CS 541: Artificial Intelligence Lecture X: Markov Decision Process Slides Credit: Peter Norvig and Sebastian Thrun."

Similar presentations


Ads by Google