Size: px
Start display at page:

Download ""

Transcription

1

2 Eligibility traces Chapter 12, plus some extra stuff! Like n-step methods, but better!

3 Eligibility traces A mechanism that allow TD, Sarsa and Q-learning to learn more efficiently A way to move smoothly between TD(0) updates and Monte Carlo updates (using full returns) - useful when we don t have the Markov property - allows us to do Monte Carlo-like updates even when the task is continuing or episodes are very long - methods in-between are often better in practice Allows TD methods to due multi-step updates More simple to implement than n-step methods

4 1-step prediction Consider the update target of contant-α MC - G t R t+1 + γr t+2 + γ 2 R t γ T-t-1 R T - uses the full return, all future rewards until termination Consider the update target of TD(0) - G (1) t R t+1 + γv(s t+1 ) - it s only based the next reward and γv(s t+1 ) is a stand-in for the rest of the return γr t+2 + γ 2 R t this allows us to avoid waiting, we can update on every step

5 n-returns prediction We do something in between - use more than one reward, but less than all rewards until termination Use two rewards and then the value function: - G (2) t R t+1 + γr t+2 + γ 2 V(S t+2 ) - now γ 2 V(S t+2 ) takes the place of γ 2 R t+3 + γ 3 R t we call this a 2-step return We could do 3-step, 4-step,.. n-steps - G (n) t R t+1 + γr t+2 + γ 2 R t γ n-1 R t+n + γ n V(S t+n ) We call these n-step TD methods; TD(0) is a one-step TD method

6 Spectrum of returns TD (1-step) 2-step 3-step n-step Monte Carlo

7 n-step returns All n-step returns are approximations of the full return Gt - truncated after n-steps and corrected for the remaining missing terms using V(St+n) Monte Carlo methods are a special case being an -step return

8 TD updates with n-step returns We can do these updates online, during the episode as soon as the information becomes available: - we wait n-1 steps We have already explored implementations of n- step methods - they work better than TD(0) - BUT the algorithms are more complex

9 Complex backups Another valid option is averaging n-step returns For example we could average a 2-step and 4-step return G (2) t G (4) t In fact, we can average any set of returns this way, even infinite set of returns as long as the weighting sums to 1 - still guaranteed to converge to correct predictions The average of simple backups is called a complex backup, TD(λ) uses a complex backup 1 2

10 λ-returns The TD(λ) algorithm averages n-step returns in a particular way Each n-step return is weighted proportional to λ n-1, where λ [0,1] For the episodic case the λ-return is: TXt 1 G t = (1 ) n=1 n 1 G (n) t + T t 1 G t, G (n) t def = R t+1 + R t R t n 1 R t+n + n V (S t+n ) = nx 1 k=0 k R t+k+1 + n V (S t+n ) When λ=1 G λ t =G t, the full return, Monte Carlo When λ=0 G λ t =G (1) t, the 1-step return, TD(0), hence the name!

11 λ-return example T t 1 G t = (1 ) Let T = 4 and t = 0. X n=1 The one step return (n = 1 of the sum): n 1 G (n) t + T t 1 G t, (1 ) (n 1) G (n) t =(1 ) (0) G (1) t =(1 )[R t+1 + V (S t+1 )] =(1 )[R 1 + V (S 1 )] The two step return (n = 2 of the sum): (1 ) (1) G (2) t =(1 ) [R 1 + R V (S 2 )] The two step return (n = 3 of the sum): (1 ) (2) G (3) t =(1 ) 2 [R 1 + R R V (S 3 )] now we stop because T t 1=4 0 1 = 3. No we take care of the last component of the equation: T t 1 G t = 3 [R 1 + R R R T =4 ]

12 given the next largest weight, (1 ) ; the three-step return is given the weight ) 2 ; and so on. The weight fades by with each additional step. After a rminal state has been reached, all subsequent n-step returns are equal to Gt. If we ant, we can separate these post-termination terms from the main sum, yielding λ-returns 7.2. THE FORWARD VIEW OF TD( ) Gt = (1 ) TX t 1 n 1 (n) Gt T t 1 + Gt, (7.6) TD("), "-return n=1 indicated in the figures. This equation makes it clearer what happens when = 1. In this case the main sum goes to zero, and the remaining term reduces to e conventional return, Gt. Thus, for = 1, backing up according to the -return def (n) Monte the same as the G = R Carlo + algorithm V (S that ) we called constant- MC (6.1) in the Weight t+1 t (n) Gt def = t+1 weight given to the 3-step return Rt+1 R 2 (1 ) t+2 is !" total area = 1 (1!") " V (St+2 ) decay by " (n) G1!" t (n) Gt def = Rt+1 + Rt+2 + def = Rt+1 + Rt+2 + t 2 Rt V (St+3 )weight given to actual, final return is T t 1 Rt T # n 1=1 Rt+n + 2 (1!") " n V (St+n ) Time Figure 7.3: The backup digram for TD( ). If T-t-1 " = 0, then the overa

13 t want, we can separate these post-termination terms from the main sum, yielding is given the next largest weight, (1 ) ; the three-step return is given the TX t 1 2 (1 ) ; and so step. A n on. 1 (n) T t fades 1 Gt = (1 ) Gt The+weight Gt, by with each additional(7.6) terminal state has been reached, all subsequent n-step returns are equal to Gt n=1 want, we can separate these post-termination terms from the main sum, yield as indicated in the figures. This equation makes it clearer what happens when t 1 = 1. In this case the main sumt X goes to zero, and thet remaining term reduces to n 1 (n) t 1 Gt = G(1. Thus, ) for = 1, Gt backing + up according Gt, the conventional return, to the -return t n=1 is the same as the Monte Carlo algorithm that we called constant- MC (6.1) in the as indicated in the figures. This equation makes it clearer what happens = 1. In this case the main sum goes to zero, and the remaining term redu weight given to the conventional return, =total 1, backing area = 1up according to the the 3-step return Gt. Thus, for (1 Monte ) 2 is the same asisthe Carlo algorithm that we called constant- MC (6.1) λ-returns decay by " Weight weight given to the 3-step return ) 2 is (1 1!" weight given total toarea = 1 actual, final return is T t 1 decay by " Weight t 1!" T Time weight given to actual, final return is T t 1 Figure 7.4: Weighting given in the -return to each of the n-step returns. T t Time

14 Can we show that the weighting sums to 1? TXt 1 G t = (1 ) n=1 n 1 G (n) t + T t 1 G t, Consider the first term: (1 ) k=a TXt 1 n=1 n 1 We can use the generalized formula for the sum of geometric series: bx r k = ra r b+1 1 r

15 TXt 1 Consider the first term: (1 ) bx k=a r k = ra r b+1 1 r n=1 n 1 generalized formula TXt 1 n=1 n = 1 T t plug in the variables. But we want λ^{n-1} not n TXt 1 n=1 n 1 = TXt 1 n=1 n ( 1 )= 1 T X t 1 n=1 n now formula in correct form TXt 1 n=1 n 1 = (1 T t 1 ) (1 ) = 1 T t 1 (1 ) pull λ out of numerator and cancel λ (1 ) TXt 1 n 1 =1 T t 1 n=1 add in the last term: 1 T t 1 + T t 1 =1

16 λ-return updates The λ-return algorithm, performs backups towards the λ-return as a target: G (λ) t t+1 h i. = t + G t ˆv(S t, t ) rˆv(s t, t ) The off-line algorithm makes no changes to the weight vector during the episode All updates are done at the end of episodes

17 19 State chain (no approximation = no FA) THE -RETURN 255 n-step TD methods Off-line λ-return algorithm λ= λ=.99 (from Chapter 7) 128 n=64 n=32 λ=.975 RMS error at the end of the episode over the first 10 episodes λ=.95 λ=0 λ=.95 n=32 n=1 n=16 λ=.9 λ=.4 λ=.8 n=8 n=2 n=4 Figure 12.3: 19-state Random walk results (Example 7.1): Performance of the o ine return algorithm alongside that of the n-step TD methods. In both case, intermediate values

18 1!" The forward view (1!") " 2 The λ-return algorithm uses what is called the forward view (1!") " #=1 " From each state we visit, we look forward in time, 7.3. THEfuture BACKWARD VIEWFigure OF TD(7.3: ) The backup digram for TD( ). If155= 0, then th to all rewards first component, the one-step TD backup, whereas if to its last component, the Monte Carlo backup. Rr = 1, th T Rrt+3 Sst+3 +3 Rrt+1 Sstt is given next largest weight, (1 ) ; the three-step Rrt+2 the Sstt (1 ) ; and so on. The weight fades by with eac Sst+1 terminal state has been reached, all subsequent n-step re want, T imewe can separate these post-termination terms from TX t 1 (n) by looking Figure 7.6: The forward or theoretical view. We decide how to update each n 1 state T t 1 G = (1 ) G + Gt, t t forward to future rewards and states. n=1

19 λ-return and Monte Carlo The λ-return algorithm uses the λ parameter to shift between TD(0) and Monte Carlo backups - same way that n in n-step TD method allowed us to shift between the two extremes The λ-return is distinctive because there is a special algorithm for achieving it, that does not require looking into the future - instead we keep a memory and make updates to previously visited states - this is what the TD(λ) algorithm does

20 update treated in Chapter 9 (and, in the tabular case, to the simple TD rule (6.2)). Thisbackward is why that algorithm was called The view looks backtd(0). In terms of Figure 12.5, TD(0) is the case in which only the one state preceding the current one is changed by the totdtheerror. recently visited states (marked<by eligibility traces) For larger values of, but still 1, more of the preceding states eett Sst-3-3 eett Sstt-2-2!t t eett Sstt-1-1 eett Sstt T ime Sstt+1 +1 Figure Shout 12.5: The or mechanistic view. Each update depends on the current TD thebackward TD error backwards error combined with eligibility traces of past events. The traces fade with temporal distance by γλ

21 The backward view The forward view, is theoretical, it s not directly implementable - at each step it requires knowledge of what will happen on future steps The backward view provides a mechanism for approximating the forward view We need a memory variable called the eligibility trace

22 The TD(λ) algorithm One of the oldest, best known, and most successful algorithms in RL Three distinctive advantages over off-line λ-return algorithm: - updates weight vector (value func) on every step, not just at the end of episodes - computation equally distributed (not just at end) - can be applied to continuing problems

23 Eligibility traces The eligibility traces keep a record of which states have been visited recently The traces indicates the degree to which a state is eligible for a learning update accumulating eligibility trace times of visits to a state

24 Demo Here we are marking state-action pairs with a replacing eligibility trace 24

25 Accumulating eligibility traces same shape as θ New memory vector called eligibility trace e t 2 R n 0 - On each step, decay each component by γλ and increment the trace for the current state by 1 e 0. = 0, e t. = rˆv(st, t )+ e t 1 accumulating eligibility trace times of visits to a state

26 Combining eligibility traces and TD We want to send information back in time - but what do we send back? The temporal difference error is defined as: t def = R t+1 + ˆv(S t+1, t ) ˆv(S t, t ) Using the eligibility trace we update the values of previously visited states with the TD error on this time step def = t e t

27 the semi-gradient version of TD( ) with function approx TD errors.function The error forrate state-value prediction is is where is thetddiscount and is the paramete With approximation, the eligibility trace section. eligibility trace trackvector of which com same number components as keeps the weight t. Wh. The of v (St, t ). t = Rt+1 + v (St+1, t ) have contributed, or negatively, to recent st long-term memory, positively accumulating over the lifetime of the istd( defined in terms. algorithm The is said th is a short-term memory, typically lasting lessindicate time TheInSemi-gradient TD(λ) ), the weight vector istrace updated on to each step than pro Eligibility traces assist in the learning process; their onl of the weight vector for undergoing learning changes sh error and the vector eligibility trace: a ect the weight vector, andare thenconcerned the weightwith vector The reinforcing events we aredete the. + eligibility, TD TD for state-value predictiontoisz t+1 =), The tthe t eterror In errors. TD( trace vector is initialized episode, is.incremented on each time step by the value gra On the next page, complete pseudocode for TD( ) is give = R + v (S, ) v (S, ). t t+1 t+1 t t t by : of its operation is suggested by Figure In TD( the vectorinis time. updated on each step 0, weight TD( )e0is),= oriented backward At each moment. et assign =the rv (S etto1,each errorand and vector eligibility trace: t) + error itt, backward prior state accordin contributed to. the current eligibility trace at that time. W where t+1 is = the t discount + t et, rate and is the parameter i section. The eligibility trace keeps track of which comp have contributed, negatively, to On the next page,positively completeorpseudocode forrecent TD( state ) is g is in terms. The trace said to12.5. indicate the e of defined its operation is suggested by is Figure

28 TD(1) The TD(1) algorithm is more general than Monte Carlo method - it can be applied to continuing tasks - can be performed incrementally and online - learns during episodes - control methods based on TD(1) can learn from unusually good or bad rewards during the episode and alter the behavior during the episode

29 TD(λ) performs similarly to offline λ-return alg. but slightly worse, particularly at high α Tabular 19-state random walk task RMS error at the end of the episode over the first 10 episodes λ=.95 TD(λ) λ=.9 λ=0 λ=.8 λ=.95 λ=1 Off-line λ-return algorithm (from the previous section) λ=.99 λ=.975 λ=.95 λ=0 λ=.9 λ=.8 λ=.4 λ=.9 λ=.8 λ=.4 Can we do better? Can we update online?

30 Eligibility in the linear case Assuming linear FA - TD(λ) with accumulating traces updates et: - et = γλet-1 + ɸt Assuming binary features (like tile coding) - et = γλet-1 - e(f)t = e(f)t where F is the indexes of the active features

31 Replacing traces The previous experiment suggests a weakness of TD(λ) with accumulating traces With accumulating traces multiple visits to a state can cause further increments to e t Instead of: e t = γλe t-1 + ɸ t, we could replace the trace value: - e t = max(γλe t-1, ɸ t ) This is called a replacing trace If λ=1 TD(λ) with replacing traces is closely related to first-visit Monte Carlo

32 Replacing traces for binary features et = γλet-1 e(f)t = 1 F contains the indexes of the active features

33 Linear True-online methods The λ-return algorithm is ideally what we would like to do It performs better than TD(λ) in practice TD(λ) is only equivalent to the λ-return algorithm at the ends of episodes: - after each episode, when each algorithm is done all their updates, they are the same Recently a new version of TD(λ) was proposed that achieves equivalence during the episode!

34 Dutch traces, linear case e t. = et e > t 1 t t This is called a dutch trace - somewhere in between accumulating and replacing traces When used with TD(λ) we call it: True-online TD(λ)

35 T role in bootstrapping in the n-step returns of the updates. In the fin. diagonal weight vectors are renamed without a superscript, t = then is to find a compact, efficient way of computing each tt from th this is done, for the linear case in which v (s, ) = > (s), then we a online TD( ) algorithm:. > > t+1 = t + t et + t t t 1 t (et t ), True online TD(λ) 268 where we have used the shorthand et is defined by. et = et > et 1 t. = t CHAPTER 12. ELIGIB (St ), t is defined as in TD t. dutch trace This algorithm has been proven to produce exactly the same seque tors, t, 80 t T, as the on-line -return algorithm (van Siejen the results on the random walk task on the left of Figure 12.7 are that task. Now, however, the algorithm is much less expensive. The ments of true online TD( ) are identical to those of conventional per-step computation is increased by about 50% (there is one more

36 36 Accumulating, Dutch, and Replacing Traces All traces fade the same: But increment differently! times of state visits accumulating traces dutch traces (α = 0.5) replacing traces

37 Problems with accumulating traces wrong wrong wrong wrong wrong right right right right +1 right Consider the above MDP: learning a policy (and q) Long sequence of states, so we would like to use long traces Imagine in the first episode, in some state s, we took the wrong action several times and then the right action - right action is more recent, but e corresponding to (s,wrong) > e corresponding to (s,right) - at termination Q(s,wrong) > Q(s,right) - which further compounds the problem Not a problem for replacing traces and Sarsa

38 Chain On-line TD(λ), replacing traces λ=1 λ=.99 On-line TD(λ), dutch traces λ=1 λ=.975 λ=.99 RMS error over first 10 episodes λ=.975 λ=.95 λ=0 λ=.95 λ=.975 λ=.95 λ=0 λ=.9 λ=.8 λ=.4 λ=.9 λ=.8 λ=.4

39 Comparison with λ-return algorithm On-line TD(λ), replacing traces λ=1 λ=.99 On-line TD(λ), dutch traces λ=1 λ=.975 λ=.99 RMS error over first 10 episodes λ=.975 λ=.95 λ=0 λ=.95 λ=.975 λ=.95 λ=0 λ=.9 λ=.8 λ=.4 λ=.9 λ=.8 λ=.4 On-line λ-return algorithm λ=1 λ=.99 λ=.975 λ=.95 Off-line λ-return algorithm off-line TD(λ), accumulating traces λ=1 λ=.99 λ=0 RMS error over first 10 episodes λ=.95 λ=0 λ=.99 λ=.975 λ=.95 λ=.9 λ=.8 λ=.4 λ=.9 λ=.8 λ=.4

40 Comparison of all three traces (online) RMS error over first 10 episodes On-line TD(λ), accumulating traces λ=.9 λ=.8 λ=.95 λ=.9 λ=.8 λ=0 λ=.4 On-line TD(λ), replacing traces λ=1 λ=.99 Off-line TD(λ), accumulating traces off-line λ-return algorithm λ=1 λ=.99 λ=0 λ=.4 λ=.975 λ=.95 λ=.8 λ=.9 λ=.99 On-line TD(λ), dutch traces λ=1 λ=.975 λ=.99 RMS error over first 10 episodes λ=.975 λ=.95 λ=0 λ=.95 λ=.975 λ=.95 λ=0 λ=.9 λ=.8 λ=.4 λ=.9 λ=.8 λ=.4

41 The winner is? Dutch traces is the closest match to the λ-return algorithm, under online updating Replacing traces gets lower error and does much better with larger λ values Replacing traces don t extend in a general way to function approximation You can create small MDPs where λ makes no difference for TD(λ) with replacing traces You can create small MDPs where large λ and large α causes TD(λ) with accumulating traces to blow up!

42 TD(λ) + Generalized policy iteration + FA As usual we combing policy evaluation for q functions with a policy improvement step TD(λ) handles the policy evaluation step And we form the policy to be ϵ-greedy with respect to the current q estimate The eligibility traces are now defined over s,a pairs

43 Linear semi-gradient Sarsa(λ) with binary features and ϵ-greedy policy Let and e be vectors with one component for each possible feature Let F a, for every possible action a, be a set of feature indices, initially empty Initialize as appropriate for the problem, e.g., = 0 Repeat (for each episode): e = 0 S, A initial state and action of episode (e.g., "-greedy) F A set of features present in S, A Repeat (for each step of episode): For all i 2 F A : e i e i + 1 (accumulating traces) or e i 1 (replacing traces) Take action A, observe reward, R, and next state, S 0 R P i2f A i If S 0 is terminal, then + e; go to next episode For all a 2 A(S 0 ): F a set P of features present in S 0,a Q a i2f a i A 0 new action in S 0 (e.g., "-greedy) + Q A 0 + e e e S S 0 A A 0

44 MOUNTAIN CAR Goal Sarsa + Tile coding on Mountain Car REPLACE TRACES 800 ACCUMULATE TRACES!= 1!= !=. 5 Steps per episode averaged over first first episodes trials and 30 runs!=. 9 9!= 0!= !=. 6!= 0!=. 4!= !=. 3!= " " 5

45 Mountain Car with Radial Basis Functions S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 32

46 Should we bootstrap? In all cases lower is better Red points are the cases of no bootstrapping Pure bootstrapping No bootstrapping

47 Multi-step updating vs onestep updates with Sarsa Path taken Action values increased by one-step Sarsa Action values increased by Sarsa(!) with!=0.9

48 Efficient trace implementations Methods with traces are more complex than onestep methods! - on every step we update e and q for every (s,a) - not a problem on a GPU or vector processor Most of the time, many et are nearly zero - decay is exponential in γλ Could keep track of et > η and only decay those & update the corresponding q(s,a) values

49 Other GPI TD algorithms that use traces Watkins Q(λ) = Q-learning + eligibility traces - et=0 when-ever μ takes an action different from π In practice this means Watkins Q(λ) is hardly more efficient than regular Q-learning - if the behavior policy explores a lot

50 Variable λ λ need not be fixed to a constant We could make it a function of state - λ t = λ(s t ) - different than a function of time If we had some way to compute our confidence in V(S t ) we could: - make λ t large in states where our confidence is low: bootstrap less, trust V less, use more of returns (MC) - make λ t near zero in states where confidence is high: full bootstrapping, like a TD(0) update

51 Summary Provide an efficient, incremental way to combine Monte Carlo (MC) and temporal-difference (TD) learning methods - Includes advantages of MC (can deal with lack of Markov property) - Includes advantages of TD (using TD error, bootstrapping) Can significantly speed learning, with small small cost in computation ( x2) Extends to control in on-policy: Sarsa(λ)) Three varieties: accumulating, replacing, and dutch True online TD(λ) is new and best - Is exactly equivalent to online λ-return algorithm

52 Unified View Temporaldifference learning width of backup Dynamic programming height (depth) of backup Monte Carlo Exhaustive search...

Temporal difference learning

Temporal difference learning Temporal difference learning AI & Agents for IET Lecturer: S Luz http://www.scss.tcd.ie/~luzs/t/cs7032/ February 4, 2014 Recall background & assumptions Environment is a finite MDP (i.e. A and S are finite).

More information

Chapter 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1

Chapter 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1 Chapter 7: Eligibility Traces R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1 Midterm Mean = 77.33 Median = 82 R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction

More information

Reinforcement Learning II

Reinforcement Learning II Reinforcement Learning II Andrea Bonarini Artificial Intelligence and Robotics Lab Department of Electronics and Information Politecnico di Milano E-mail: bonarini@elet.polimi.it URL:http://www.dei.polimi.it/people/bonarini

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Temporal Difference Learning Temporal difference learning, TD prediction, Q-learning, elibigility traces. (many slides from Marc Toussaint) Vien Ngo MLR, University of Stuttgart

More information

Reinforcement Learning. Spring 2018 Defining MDPs, Planning

Reinforcement Learning. Spring 2018 Defining MDPs, Planning Reinforcement Learning Spring 2018 Defining MDPs, Planning understandability 0 Slide 10 time You are here Markov Process Where you will go depends only on where you are Markov Process: Information state

More information

Temporal Difference. Learning KENNETH TRAN. Principal Research Engineer, MSR AI

Temporal Difference. Learning KENNETH TRAN. Principal Research Engineer, MSR AI Temporal Difference Learning KENNETH TRAN Principal Research Engineer, MSR AI Temporal Difference Learning Policy Evaluation Intro to model-free learning Monte Carlo Learning Temporal Difference Learning

More information

Machine Learning I Reinforcement Learning

Machine Learning I Reinforcement Learning Machine Learning I Reinforcement Learning Thomas Rückstieß Technische Universität München December 17/18, 2009 Literature Book: Reinforcement Learning: An Introduction Sutton & Barto (free online version:

More information

Reinforcement Learning. George Konidaris

Reinforcement Learning. George Konidaris Reinforcement Learning George Konidaris gdk@cs.brown.edu Fall 2017 Machine Learning Subfield of AI concerned with learning from data. Broadly, using: Experience To Improve Performance On Some Task (Tom

More information

Reinforcement Learning. Value Function Updates

Reinforcement Learning. Value Function Updates Reinforcement Learning Value Function Updates Manfred Huber 2014 1 Value Function Updates Different methods for updating the value function Dynamic programming Simple Monte Carlo Temporal differencing

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Function approximation Daniel Hennes 19.06.2017 University Stuttgart - IPVS - Machine Learning & Robotics 1 Today Eligibility traces n-step TD returns Forward and backward view Function

More information

Reinforcement Learning. Summer 2017 Defining MDPs, Planning

Reinforcement Learning. Summer 2017 Defining MDPs, Planning Reinforcement Learning Summer 2017 Defining MDPs, Planning understandability 0 Slide 10 time You are here Markov Process Where you will go depends only on where you are Markov Process: Information state

More information

The Book: Where we are and where we re going. CSE 190: Reinforcement Learning: An Introduction. Chapter 7: Eligibility Traces. Simple Monte Carlo

The Book: Where we are and where we re going. CSE 190: Reinforcement Learning: An Introduction. Chapter 7: Eligibility Traces. Simple Monte Carlo CSE 190: Reinforcement Learning: An Introduction Chapter 7: Eligibility races Acknowledgment: A good number of these slides are cribbed from Rich Sutton he Book: Where we are and where we re going Part

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Temporal Difference Learning Temporal difference learning, TD prediction, Q-learning, elibigility traces. (many slides from Marc Toussaint) Vien Ngo Marc Toussaint University of

More information

Review: TD-Learning. TD (SARSA) Learning for Q-values. Bellman Equations for Q-values. P (s, a, s )[R(s, a, s )+ Q (s, (s ))]

Review: TD-Learning. TD (SARSA) Learning for Q-values. Bellman Equations for Q-values. P (s, a, s )[R(s, a, s )+ Q (s, (s ))] Review: TD-Learning function TD-Learning(mdp) returns a policy Class #: Reinforcement Learning, II 8s S, U(s) =0 set start-state s s 0 choose action a, using -greedy policy based on U(s) U(s) U(s)+ [r

More information

PART A and ONE question from PART B; or ONE question from PART A and TWO questions from PART B.

PART A and ONE question from PART B; or ONE question from PART A and TWO questions from PART B. Advanced Topics in Machine Learning, GI13, 2010/11 Advanced Topics in Machine Learning, GI13, 2010/11 Answer any THREE questions. Each question is worth 20 marks. Use separate answer books Answer any THREE

More information

arxiv: v1 [cs.ai] 5 Nov 2017

arxiv: v1 [cs.ai] 5 Nov 2017 arxiv:1711.01569v1 [cs.ai] 5 Nov 2017 Markus Dumke Department of Statistics Ludwig-Maximilians-Universität München markus.dumke@campus.lmu.de Abstract Temporal-difference (TD) learning is an important

More information

Lecture 23: Reinforcement Learning

Lecture 23: Reinforcement Learning Lecture 23: Reinforcement Learning MDPs revisited Model-based learning Monte Carlo value function estimation Temporal-difference (TD) learning Exploration November 23, 2006 1 COMP-424 Lecture 23 Recall:

More information

Reinforcement Learning: An Introduction. ****Draft****

Reinforcement Learning: An Introduction. ****Draft**** i Reinforcement Learning: An Introduction Second edition, in progress ****Draft**** Richard S. Sutton and Andrew G. Barto c 2014, 2015 A Bradford Book The MIT Press Cambridge, Massachusetts London, England

More information

Grundlagen der Künstlichen Intelligenz

Grundlagen der Künstlichen Intelligenz Grundlagen der Künstlichen Intelligenz Reinforcement learning II Daniel Hennes 11.12.2017 (WS 2017/18) University Stuttgart - IPVS - Machine Learning & Robotics 1 Today Eligibility traces n-step TD returns

More information

Replacing eligibility trace for action-value learning with function approximation

Replacing eligibility trace for action-value learning with function approximation Replacing eligibility trace for action-value learning with function approximation Kary FRÄMLING Helsinki University of Technology PL 5500, FI-02015 TKK - Finland Abstract. The eligibility trace is one

More information

This question has three parts, each of which can be answered concisely, but be prepared to explain and justify your concise answer.

This question has three parts, each of which can be answered concisely, but be prepared to explain and justify your concise answer. This question has three parts, each of which can be answered concisely, but be prepared to explain and justify your concise answer. 1. Suppose you have a policy and its action-value function, q, then you

More information

Least squares temporal difference learning

Least squares temporal difference learning Least squares temporal difference learning TD(λ) Good properties of TD Easy to implement, traces achieve the forward view Linear complexity = fast Combines easily with linear function approximation Outperforms

More information

Open Theoretical Questions in Reinforcement Learning

Open Theoretical Questions in Reinforcement Learning Open Theoretical Questions in Reinforcement Learning Richard S. Sutton AT&T Labs, Florham Park, NJ 07932, USA, sutton@research.att.com, www.cs.umass.edu/~rich Reinforcement learning (RL) concerns the problem

More information

Lecture 7: Value Function Approximation

Lecture 7: Value Function Approximation Lecture 7: Value Function Approximation Joseph Modayil Outline 1 Introduction 2 3 Batch Methods Introduction Large-Scale Reinforcement Learning Reinforcement learning can be used to solve large problems,

More information

Week 6, Lecture 1. Reinforcement Learning (part 3) Announcements: HW 3 due on 11/5 at NOON Midterm Exam 11/8 Project draft due 11/15

Week 6, Lecture 1. Reinforcement Learning (part 3) Announcements: HW 3 due on 11/5 at NOON Midterm Exam 11/8 Project draft due 11/15 ME 537: Learning Based Control Week 6, Lecture 1 Reinforcement Learning (part 3 Announcements: HW 3 due on 11/5 at NOON Midterm Exam 11/8 Project draft due 11/15 Suggested reading : Chapters 7-8 in Reinforcement

More information

Generalization and Function Approximation

Generalization and Function Approximation Generalization and Function Approximation 0 Generalization and Function Approximation Suggested reading: Chapter 8 in R. S. Sutton, A. G. Barto: Reinforcement Learning: An Introduction MIT Press, 1998.

More information

CSE 190: Reinforcement Learning: An Introduction. Chapter 8: Generalization and Function Approximation. Pop Quiz: What Function Are We Approximating?

CSE 190: Reinforcement Learning: An Introduction. Chapter 8: Generalization and Function Approximation. Pop Quiz: What Function Are We Approximating? CSE 190: Reinforcement Learning: An Introduction Chapter 8: Generalization and Function Approximation Objectives of this chapter: Look at how experience with a limited part of the state set be used to

More information

Basics of reinforcement learning

Basics of reinforcement learning Basics of reinforcement learning Lucian Buşoniu TMLSS, 20 July 2018 Main idea of reinforcement learning (RL) Learn a sequential decision policy to optimize the cumulative performance of an unknown system

More information

Introduction to Reinforcement Learning. CMPT 882 Mar. 18

Introduction to Reinforcement Learning. CMPT 882 Mar. 18 Introduction to Reinforcement Learning CMPT 882 Mar. 18 Outline for the week Basic ideas in RL Value functions and value iteration Policy evaluation and policy improvement Model-free RL Monte-Carlo and

More information

Reinforcement Learning with Function Approximation. Joseph Christian G. Noel

Reinforcement Learning with Function Approximation. Joseph Christian G. Noel Reinforcement Learning with Function Approximation Joseph Christian G. Noel November 2011 Abstract Reinforcement learning (RL) is a key problem in the field of Artificial Intelligence. The main goal is

More information

ROB 537: Learning-Based Control. Announcements: Project background due Today. HW 3 Due on 10/30 Midterm Exam on 11/6.

ROB 537: Learning-Based Control. Announcements: Project background due Today. HW 3 Due on 10/30 Midterm Exam on 11/6. ROB 537: Learning-Based Control Week 5, Lecture 1 Policy Gradient, Eligibility Traces, Transfer Learning (MaC Taylor Announcements: Project background due Today HW 3 Due on 10/30 Midterm Exam on 11/6 Reading:

More information

PART A and ONE question from PART B; or ONE question from PART A and TWO questions from PART B.

PART A and ONE question from PART B; or ONE question from PART A and TWO questions from PART B. Advanced Topics in Machine Learning, GI13, 2010/11 Advanced Topics in Machine Learning, GI13, 2010/11 Answer any THREE questions. Each question is worth 20 marks. Use separate answer books Answer any THREE

More information

CS599 Lecture 2 Function Approximation in RL

CS599 Lecture 2 Function Approximation in RL CS599 Lecture 2 Function Approximation in RL Look at how experience with a limited part of the state set be used to produce good behavior over a much larger part. Overview of function approximation (FA)

More information

Chapter 6: Temporal Difference Learning

Chapter 6: Temporal Difference Learning Chapter 6: emporal Difference Learning Objectives of this chapter: Introduce emporal Difference (D) learning Focus first on policy evaluation, or prediction, methods Compare efficiency of D learning with

More information

CS599 Lecture 1 Introduction To RL

CS599 Lecture 1 Introduction To RL CS599 Lecture 1 Introduction To RL Reinforcement Learning Introduction Learning from rewards Policies Value Functions Rewards Models of the Environment Exploitation vs. Exploration Dynamic Programming

More information

CSE 190: Reinforcement Learning: An Introduction. Chapter 8: Generalization and Function Approximation. Pop Quiz: What Function Are We Approximating?

CSE 190: Reinforcement Learning: An Introduction. Chapter 8: Generalization and Function Approximation. Pop Quiz: What Function Are We Approximating? CSE 190: Reinforcement Learning: An Introduction Chapter 8: Generalization and Function Approximation Objectives of this chapter: Look at how experience with a limited part of the state set be used to

More information

Lecture 3: Policy Evaluation Without Knowing How the World Works / Model Free Policy Evaluation

Lecture 3: Policy Evaluation Without Knowing How the World Works / Model Free Policy Evaluation Lecture 3: Policy Evaluation Without Knowing How the World Works / Model Free Policy Evaluation CS234: RL Emma Brunskill Winter 2018 Material builds on structure from David SIlver s Lecture 4: Model-Free

More information

Notes on Reinforcement Learning

Notes on Reinforcement Learning 1 Introduction Notes on Reinforcement Learning Paulo Eduardo Rauber 2014 Reinforcement learning is the study of agents that act in an environment with the goal of maximizing cumulative reward signals.

More information

Machine Learning. Reinforcement learning. Hamid Beigy. Sharif University of Technology. Fall 1396

Machine Learning. Reinforcement learning. Hamid Beigy. Sharif University of Technology. Fall 1396 Machine Learning Reinforcement learning Hamid Beigy Sharif University of Technology Fall 1396 Hamid Beigy (Sharif University of Technology) Machine Learning Fall 1396 1 / 32 Table of contents 1 Introduction

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning RL in continuous MDPs March April, 2015 Large/Continuous MDPs Large/Continuous state space Tabular representation cannot be used Large/Continuous action space Maximization over action

More information

Chapter 8: Generalization and Function Approximation

Chapter 8: Generalization and Function Approximation Chapter 8: Generalization and Function Approximation Objectives of this chapter: Look at how experience with a limited part of the state set be used to produce good behavior over a much larger part. Overview

More information

15-780: Graduate Artificial Intelligence. Reinforcement learning (RL)

15-780: Graduate Artificial Intelligence. Reinforcement learning (RL) 15-780: Graduate Artificial Intelligence Reinforcement learning (RL) From MDPs to RL We still use the same Markov model with rewards and actions But there are a few differences: 1. We do not assume we

More information

Prof. Dr. Ann Nowé. Artificial Intelligence Lab ai.vub.ac.be

Prof. Dr. Ann Nowé. Artificial Intelligence Lab ai.vub.ac.be REINFORCEMENT LEARNING AN INTRODUCTION Prof. Dr. Ann Nowé Artificial Intelligence Lab ai.vub.ac.be REINFORCEMENT LEARNING WHAT IS IT? What is it? Learning from interaction Learning about, from, and while

More information

ilstd: Eligibility Traces and Convergence Analysis

ilstd: Eligibility Traces and Convergence Analysis ilstd: Eligibility Traces and Convergence Analysis Alborz Geramifard Michael Bowling Martin Zinkevich Richard S. Sutton Department of Computing Science University of Alberta Edmonton, Alberta {alborz,bowling,maz,sutton}@cs.ualberta.ca

More information

What we learned last time

What we learned last time Wat we learned last time Value-function approximation by stocastic gradient descent enables RL to be applied to arbitrarily large state spaces Most algoritms just carry over Targets from tabular case Wit

More information

Reinforcement learning an introduction

Reinforcement learning an introduction Reinforcement learning an introduction Prof. Dr. Ann Nowé Computational Modeling Group AIlab ai.vub.ac.be November 2013 Reinforcement Learning What is it? Learning from interaction Learning about, from,

More information

Reinforcement Learning: An Introduction

Reinforcement Learning: An Introduction Introduction Betreuer: Freek Stulp Hauptseminar Intelligente Autonome Systeme (WiSe 04/05) Forschungs- und Lehreinheit Informatik IX Technische Universität München November 24, 2004 Introduction What is

More information

arxiv: v1 [cs.ai] 1 Jul 2015

arxiv: v1 [cs.ai] 1 Jul 2015 arxiv:507.00353v [cs.ai] Jul 205 Harm van Seijen harm.vanseijen@ualberta.ca A. Rupam Mahmood ashique@ualberta.ca Patrick M. Pilarski patrick.pilarski@ualberta.ca Richard S. Sutton sutton@cs.ualberta.ca

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Function approximation Mario Martin CS-UPC May 18, 2018 Mario Martin (CS-UPC) Reinforcement Learning May 18, 2018 / 65 Recap Algorithms: MonteCarlo methods for Policy Evaluation

More information

Lecture 8: Policy Gradient

Lecture 8: Policy Gradient Lecture 8: Policy Gradient Hado van Hasselt Outline 1 Introduction 2 Finite Difference Policy Gradient 3 Monte-Carlo Policy Gradient 4 Actor-Critic Policy Gradient Introduction Vapnik s rule Never solve

More information

Marks. bonus points. } Assignment 1: Should be out this weekend. } Mid-term: Before the last lecture. } Mid-term deferred exam:

Marks. bonus points. } Assignment 1: Should be out this weekend. } Mid-term: Before the last lecture. } Mid-term deferred exam: Marks } Assignment 1: Should be out this weekend } All are marked, I m trying to tally them and perhaps add bonus points } Mid-term: Before the last lecture } Mid-term deferred exam: } This Saturday, 9am-10.30am,

More information

Monte Carlo is important in practice. CSE 190: Reinforcement Learning: An Introduction. Chapter 6: Temporal Difference Learning.

Monte Carlo is important in practice. CSE 190: Reinforcement Learning: An Introduction. Chapter 6: Temporal Difference Learning. Monte Carlo is important in practice CSE 190: Reinforcement Learning: An Introduction Chapter 6: emporal Difference Learning When there are just a few possibilitieo value, out of a large state space, Monte

More information

CMU Lecture 12: Reinforcement Learning. Teacher: Gianni A. Di Caro

CMU Lecture 12: Reinforcement Learning. Teacher: Gianni A. Di Caro CMU 15-781 Lecture 12: Reinforcement Learning Teacher: Gianni A. Di Caro REINFORCEMENT LEARNING Transition Model? State Action Reward model? Agent Goal: Maximize expected sum of future rewards 2 MDP PLANNING

More information

Off-Policy Actor-Critic

Off-Policy Actor-Critic Off-Policy Actor-Critic Ludovic Trottier Laval University July 25 2012 Ludovic Trottier (DAMAS Laboratory) Off-Policy Actor-Critic July 25 2012 1 / 34 Table of Contents 1 Reinforcement Learning Theory

More information

Balancing and Control of a Freely-Swinging Pendulum Using a Model-Free Reinforcement Learning Algorithm

Balancing and Control of a Freely-Swinging Pendulum Using a Model-Free Reinforcement Learning Algorithm Balancing and Control of a Freely-Swinging Pendulum Using a Model-Free Reinforcement Learning Algorithm Michail G. Lagoudakis Department of Computer Science Duke University Durham, NC 2778 mgl@cs.duke.edu

More information

Introduction. Reinforcement Learning

Introduction. Reinforcement Learning Introduction Reinforcement Learning Learn Act Sense Scott Sanner NICTA / ANU First.Last@nicta.com.au Lecture Goals 1) To understand formal models for decisionmaking under uncertainty and their properties

More information

15-889e Policy Search: Gradient Methods Emma Brunskill. All slides from David Silver (with EB adding minor modificafons), unless otherwise noted

15-889e Policy Search: Gradient Methods Emma Brunskill. All slides from David Silver (with EB adding minor modificafons), unless otherwise noted 15-889e Policy Search: Gradient Methods Emma Brunskill All slides from David Silver (with EB adding minor modificafons), unless otherwise noted Outline 1 Introduction 2 Finite Difference Policy Gradient

More information

Reinforcement Learning Part 2

Reinforcement Learning Part 2 Reinforcement Learning Part 2 Dipendra Misra Cornell University dkm@cs.cornell.edu https://dipendramisra.wordpress.com/ From previous tutorial Reinforcement Learning Exploration No supervision Agent-Reward-Environment

More information

Forward Actor-Critic for Nonlinear Function Approximation in Reinforcement Learning

Forward Actor-Critic for Nonlinear Function Approximation in Reinforcement Learning Forward Actor-Critic for Nonlinear Function Approximation in Reinforcement Learning Vivek Veeriah Dept. of Computing Science University of Alberta Edmonton, Canada vivekveeriah@ualberta.ca Harm van Seijen

More information

Grundlagen der Künstlichen Intelligenz

Grundlagen der Künstlichen Intelligenz Grundlagen der Künstlichen Intelligenz Reinforcement learning Daniel Hennes 4.12.2017 (WS 2017/18) University Stuttgart - IPVS - Machine Learning & Robotics 1 Today Reinforcement learning Model based and

More information

Lecture 25: Learning 4. Victor R. Lesser. CMPSCI 683 Fall 2010

Lecture 25: Learning 4. Victor R. Lesser. CMPSCI 683 Fall 2010 Lecture 25: Learning 4 Victor R. Lesser CMPSCI 683 Fall 2010 Final Exam Information Final EXAM on Th 12/16 at 4:00pm in Lederle Grad Res Ctr Rm A301 2 Hours but obviously you can leave early! Open Book

More information

Reinforcement Learning. Machine Learning, Fall 2010

Reinforcement Learning. Machine Learning, Fall 2010 Reinforcement Learning Machine Learning, Fall 2010 1 Administrativia This week: finish RL, most likely start graphical models LA2: due on Thursday LA3: comes out on Thursday TA Office hours: Today 1:30-2:30

More information

Introduction to Reinforcement Learning. Part 5: Temporal-Difference Learning

Introduction to Reinforcement Learning. Part 5: Temporal-Difference Learning Introduction to Reinforcement Learning Part 5: emporal-difference Learning What everybody should know about emporal-difference (D) learning Used to learn value functions without human input Learns a guess

More information

6 Reinforcement Learning

6 Reinforcement Learning 6 Reinforcement Learning As discussed above, a basic form of supervised learning is function approximation, relating input vectors to output vectors, or, more generally, finding density functions p(y,

More information

Chapter 6: Temporal Difference Learning

Chapter 6: Temporal Difference Learning Chapter 6: emporal Difference Learning Objectives of this chapter: Introduce emporal Difference (D) learning Focus first on policy evaluation, or prediction, methods hen extend to control methods R. S.

More information

Introduction to Reinforcement Learning. Part 6: Core Theory II: Bellman Equations and Dynamic Programming

Introduction to Reinforcement Learning. Part 6: Core Theory II: Bellman Equations and Dynamic Programming Introduction to Reinforcement Learning Part 6: Core Theory II: Bellman Equations and Dynamic Programming Bellman Equations Recursive relationships among values that can be used to compute values The tree

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Ron Parr CompSci 7 Department of Computer Science Duke University With thanks to Kris Hauser for some content RL Highlights Everybody likes to learn from experience Use ML techniques

More information

CS 287: Advanced Robotics Fall Lecture 14: Reinforcement Learning with Function Approximation and TD Gammon case study

CS 287: Advanced Robotics Fall Lecture 14: Reinforcement Learning with Function Approximation and TD Gammon case study CS 287: Advanced Robotics Fall 2009 Lecture 14: Reinforcement Learning with Function Approximation and TD Gammon case study Pieter Abbeel UC Berkeley EECS Assignment #1 Roll-out: nice example paper: X.

More information

Lecture 3: Markov Decision Processes

Lecture 3: Markov Decision Processes Lecture 3: Markov Decision Processes Joseph Modayil 1 Markov Processes 2 Markov Reward Processes 3 Markov Decision Processes 4 Extensions to MDPs Markov Processes Introduction Introduction to MDPs Markov

More information

Christopher Watkins and Peter Dayan. Noga Zaslavsky. The Hebrew University of Jerusalem Advanced Seminar in Deep Learning (67679) November 1, 2015

Christopher Watkins and Peter Dayan. Noga Zaslavsky. The Hebrew University of Jerusalem Advanced Seminar in Deep Learning (67679) November 1, 2015 Q-Learning Christopher Watkins and Peter Dayan Noga Zaslavsky The Hebrew University of Jerusalem Advanced Seminar in Deep Learning (67679) November 1, 2015 Noga Zaslavsky Q-Learning (Watkins & Dayan, 1992)

More information

ECE276B: Planning & Learning in Robotics Lecture 16: Model-free Control

ECE276B: Planning & Learning in Robotics Lecture 16: Model-free Control ECE276B: Planning & Learning in Robotics Lecture 16: Model-free Control Lecturer: Nikolay Atanasov: natanasov@ucsd.edu Teaching Assistants: Tianyu Wang: tiw161@eng.ucsd.edu Yongxi Lu: yol070@eng.ucsd.edu

More information

MARKOV DECISION PROCESSES (MDP) AND REINFORCEMENT LEARNING (RL) Versione originale delle slide fornita dal Prof. Francesco Lo Presti

MARKOV DECISION PROCESSES (MDP) AND REINFORCEMENT LEARNING (RL) Versione originale delle slide fornita dal Prof. Francesco Lo Presti 1 MARKOV DECISION PROCESSES (MDP) AND REINFORCEMENT LEARNING (RL) Versione originale delle slide fornita dal Prof. Francesco Lo Presti Historical background 2 Original motivation: animal learning Early

More information

Linear Least-squares Dyna-style Planning

Linear Least-squares Dyna-style Planning Linear Least-squares Dyna-style Planning Hengshuai Yao Department of Computing Science University of Alberta Edmonton, AB, Canada T6G2E8 hengshua@cs.ualberta.ca Abstract World model is very important for

More information

, and rewards and transition matrices as shown below:

, and rewards and transition matrices as shown below: CSE 50a. Assignment 7 Out: Tue Nov Due: Thu Dec Reading: Sutton & Barto, Chapters -. 7. Policy improvement Consider the Markov decision process (MDP) with two states s {0, }, two actions a {0, }, discount

More information

Dual Temporal Difference Learning

Dual Temporal Difference Learning Dual Temporal Difference Learning Min Yang Dept. of Computing Science University of Alberta Edmonton, Alberta Canada T6G 2E8 Yuxi Li Dept. of Computing Science University of Alberta Edmonton, Alberta Canada

More information

Reinforcement Learning: the basics

Reinforcement Learning: the basics Reinforcement Learning: the basics Olivier Sigaud Université Pierre et Marie Curie, PARIS 6 http://people.isir.upmc.fr/sigaud August 6, 2012 1 / 46 Introduction Action selection/planning Learning by trial-and-error

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Function Approximation Continuous state/action space, mean-square error, gradient temporal difference learning, least-square temporal difference, least squares policy iteration Vien

More information

Reinforcement Learning and NLP

Reinforcement Learning and NLP 1 Reinforcement Learning and NLP Kapil Thadani kapil@cs.columbia.edu RESEARCH Outline 2 Model-free RL Markov decision processes (MDPs) Derivative-free optimization Policy gradients Variance reduction Value

More information

ARTIFICIAL INTELLIGENCE. Reinforcement learning

ARTIFICIAL INTELLIGENCE. Reinforcement learning INFOB2KI 2018-2019 Utrecht University The Netherlands ARTIFICIAL INTELLIGENCE Reinforcement learning Lecturer: Silja Renooij These slides are part of the INFOB2KI Course Notes available from www.cs.uu.nl/docs/vakken/b2ki/schema.html

More information

Reinforcement learning

Reinforcement learning Reinforcement learning Based on [Kaelbling et al., 1996, Bertsekas, 2000] Bert Kappen Reinforcement learning Reinforcement learning is the problem faced by an agent that must learn behavior through trial-and-error

More information

Internet Monetization

Internet Monetization Internet Monetization March May, 2013 Discrete time Finite A decision process (MDP) is reward process with decisions. It models an environment in which all states are and time is divided into stages. Definition

More information

The convergence limit of the temporal difference learning

The convergence limit of the temporal difference learning The convergence limit of the temporal difference learning Ryosuke Nomura the University of Tokyo September 3, 2013 1 Outline Reinforcement Learning Convergence limit Construction of the feature vector

More information

RL 3: Reinforcement Learning

RL 3: Reinforcement Learning RL 3: Reinforcement Learning Q-Learning Michael Herrmann University of Edinburgh, School of Informatics 20/01/2015 Last time: Multi-Armed Bandits (10 Points to remember) MAB applications do exist (e.g.

More information

Emphatic Temporal-Difference Learning

Emphatic Temporal-Difference Learning Emphatic Temporal-Difference Learning A Rupam Mahmood Huizhen Yu Martha White Richard S Sutton Reinforcement Learning and Artificial Intelligence Laboratory Department of Computing Science, University

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Dipendra Misra Cornell University dkm@cs.cornell.edu https://dipendramisra.wordpress.com/ Task Grasp the green cup. Output: Sequence of controller actions Setup from Lenz et. al.

More information

On and Off-Policy Relational Reinforcement Learning

On and Off-Policy Relational Reinforcement Learning On and Off-Policy Relational Reinforcement Learning Christophe Rodrigues, Pierre Gérard, and Céline Rouveirol LIPN, UMR CNRS 73, Institut Galilée - Université Paris-Nord first.last@lipn.univ-paris13.fr

More information

Lecture 9: Policy Gradient II 1

Lecture 9: Policy Gradient II 1 Lecture 9: Policy Gradient II 1 Emma Brunskill CS234 Reinforcement Learning. Winter 2019 Additional reading: Sutton and Barto 2018 Chp. 13 1 With many slides from or derived from David Silver and John

More information

Reinforcement Learning

Reinforcement Learning Reinforcement Learning Markov decision process & Dynamic programming Evaluative feedback, value function, Bellman equation, optimality, Markov property, Markov decision process, dynamic programming, value

More information

The Nature of Learning - A study of Reinforcement Learning Methodology

The Nature of Learning - A study of Reinforcement Learning Methodology The Nature of Learning - A study of Reinforcement Learning Methodology Anders Christian Kølle 1 Department of Computer Science University of Copenhagen November 21, 23 1 andersk@diku.dk One needs to explore

More information

An Introduction to Reinforcement Learning

An Introduction to Reinforcement Learning 1 / 58 An Introduction to Reinforcement Learning Lecture 01: Introduction Dr. Johannes A. Stork School of Computer Science and Communication KTH Royal Institute of Technology January 19, 2017 2 / 58 ../fig/reward-00.jpg

More information

INF 5860 Machine learning for image classification. Lecture 14: Reinforcement learning May 9, 2018

INF 5860 Machine learning for image classification. Lecture 14: Reinforcement learning May 9, 2018 Machine learning for image classification Lecture 14: Reinforcement learning May 9, 2018 Page 3 Outline Motivation Introduction to reinforcement learning (RL) Value function based methods (Q-learning)

More information

Reinforcement Learning In Continuous Time and Space

Reinforcement Learning In Continuous Time and Space Reinforcement Learning In Continuous Time and Space presentation of paper by Kenji Doya Leszek Rybicki lrybicki@mat.umk.pl 18.07.2008 Leszek Rybicki lrybicki@mat.umk.pl Reinforcement Learning In Continuous

More information

RL Lecture 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1

RL Lecture 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1 RL Lecure 7: Eligibiliy Traces R. S. Suon and A. G. Baro: Reinforcemen Learning: An Inroducion 1 N-sep TD Predicion Idea: Look farher ino he fuure when you do TD backup (1, 2, 3,, n seps) R. S. Suon and

More information

Planning in Markov Decision Processes

Planning in Markov Decision Processes Carnegie Mellon School of Computer Science Deep Reinforcement Learning and Control Planning in Markov Decision Processes Lecture 3, CMU 10703 Katerina Fragkiadaki Markov Decision Process (MDP) A Markov

More information

Chapter 3: The Reinforcement Learning Problem

Chapter 3: The Reinforcement Learning Problem Chapter 3: The Reinforcement Learning Problem Objectives of this chapter: describe the RL problem we will be studying for the remainder of the course present idealized form of the RL problem for which

More information

Markov Decision Processes and Solving Finite Problems. February 8, 2017

Markov Decision Processes and Solving Finite Problems. February 8, 2017 Markov Decision Processes and Solving Finite Problems February 8, 2017 Overview of Upcoming Lectures Feb 8: Markov decision processes, value iteration, policy iteration Feb 13: Policy gradients Feb 15:

More information

Reinforcement Learning. Yishay Mansour Tel-Aviv University

Reinforcement Learning. Yishay Mansour Tel-Aviv University Reinforcement Learning Yishay Mansour Tel-Aviv University 1 Reinforcement Learning: Course Information Classes: Wednesday Lecture 10-13 Yishay Mansour Recitations:14-15/15-16 Eliya Nachmani Adam Polyak

More information

Decision Theory: Q-Learning

Decision Theory: Q-Learning Decision Theory: Q-Learning CPSC 322 Decision Theory 5 Textbook 12.5 Decision Theory: Q-Learning CPSC 322 Decision Theory 5, Slide 1 Lecture Overview 1 Recap 2 Asynchronous Value Iteration 3 Q-Learning

More information

The Reinforcement Learning Problem

The Reinforcement Learning Problem The Reinforcement Learning Problem Slides based on the book Reinforcement Learning by Sutton and Barto Formalizing Reinforcement Learning Formally, the agent and environment interact at each of a sequence

More information

Off-Policy Temporal-Difference Learning with Function Approximation

Off-Policy Temporal-Difference Learning with Function Approximation This is a digitally remastered version of a paper published in the Proceedings of the 18th International Conference on Machine Learning (2001, pp. 417 424. This version should be crisper but otherwise

More information