- Reinforcement learning
Inspired by related psychological theory, in
computer science, reinforcement learning is a sub-area of machine learningconcerned with how an "agent" ought to take "actions" in an "environment" so as to maximize some notion of long-term "reward". Reinforcement learning algorithms attempt to find a "policy" that maps "states" of the world to the actions the agent ought to take in those states. In economicsand game theory, reinforcement learning is considered as a boundedly rational interpretation of how equilibriummay arise.
The environment is typically formulated as a finite-state
Markov decision process(MDP), and reinforcement learning algorithms for this context are highly related to dynamic programmingtechniques. State transition probabilities and reward probabilities in the MDP are typically stochastic but stationary over the course of the problem.
Reinforcement learning differs from the
supervised learningproblem in that correct input/output pairs are never presented, nor sub-optimal actions explicitly corrected. Further, there is a focus on on-line performance, which involves finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge). The exploration vs. exploitation trade-off in reinforcement learning has been mostly studied through the multi-armed banditproblem.
Formally, the basic reinforcement learning model consists of:
# a set of environment states ;
# a set of actions ; and
# a set of scalar "rewards" in .
At each time , the agent perceives its state and the set of possible actions . It chooses an action and receives from the environment the new state and a reward . Based on these interactions, the reinforcement learning agent must develop a policy which maximizes the quantity for MDPs which have a terminal state, or the quantity for MDPs without terminal states (where is some "future reward" discounting factor).
Thus, reinforcement learning is particularly well suited to problems which include a long-term versus short-term reward trade-off. It has been applied successfully to various problems, including
robot control, elevator scheduling, telecommunications, backgammonand chess(Sutton 1998, Chapter 11).
After we have defined an appropriate return function to be maximized, we need to specify the algorithm that will be used to find the policy with the maximum return.
The naive brute force approach entails the following two steps: a) For each possible policy, sample returns while following it. b) Choose the policy with the largest expected return. One problem with this is that the number of policies can be extremely large, or even infinite. Another is that returns might be stochastic, in which case a large number of samples will be required to accurately estimate the return of each policy. These problems can be ameliorated if we assume some structure and perhaps allow samples generated from one policy to influence the estimates made for another. The two main approaches for achieving this are value function estimation and direct policy optimization.
Value function approaches do this by only maintaining a set of estimates of expected returns for one policy (usually either the current or the optimal one).In such approaches one attempts to estimate either the expected return starting from state and following thereafter,
or the expected return when taking action in state and following ; thereafter,
If someone gives us for the optimal policy, we can always choose optimal actions by simply choosing the action with the highest value at each state. In order to do this using , we must either have a model of the environment, in the form of probabilities , which allow us to calculate simply through
or we can employ so-called Actor-Critic methods, in which the model is split into two parts: the critic, which maintains the state value estimate , and the actor, which is responsible for choosing the appropriate actions at each state.
Given a fixed policy , Estimating for is trivial, as one only has to average the immediate rewards. The most obvious way to do this for is to average the total return after each state. However this type of Monte Carlo sampling requires the MDP to terminate.
Thus carrying out this estimation for in the general does not seem obvious. In fact, it is quite simple once one realises that the expectation of forms a recursive
By replacing those expectations with our estimates, , and performing
gradient descentwith a squared error cost function, we obtain the temporal difference learningalgorithm TD(0). In the simplest case, the set of states and actions are both discrete and we maintain tabular estimates for each state. Similar state-action pair methods are Adaptive Heuristic Critic(AHC), SARSAand Q-Learning. All methods feature extensions whereby some approximating architecture is used, though in some cases convergenceis not guaranteed. The estimates are usually updated with some form of gradient descent, though there have been recent developments with least squaresmethods for the linear approximation case.
The above methods not only all converge to the correct estimates for a fixed policy, but can also be used to find the optimal policy. This is usually done by following a policy π that is somehow derived from the current value estimates, i.e. by choosing the action with the highest evaluation most of the time, while still occasionally taking random actions in order to explore the space. Proofs for convergence to the optimal policy also exist for the algorithms mentioned above, under certain conditions. However, all those proofs only demonstrate asymptotic convergence and little is known theoretically about the behaviour of RL algorithms in the small-sample case, apart from within very restricted settings.
An alternative method to find the optimal policy is to search directly in policy space. Policy space methods define the policy as a parameterised function with parameters . Commonly, a gradient method is employed to adjust the parameters. However, the application of gradient methods is not trivial, since no gradient information is assumed. Rather, the gradient itself must be estimated from noisy samples of the return. Since this greatly increases the computational cost, it can be advantageous to use a more powerful gradient method than steepest gradient descent. Policy space gradient methods have received a lot of attention in the last 5 years and have now reached a relatively mature stage, but they remain an active field. There are many other approaches, such as
simulated annealing, that can be taken to explore the policy space. Other direct optimization techniques, such as evolutionary computationare used in evolutionary robotics.
Current research topics include: Alternative representations (such as the
Predictive State Representationapproach), gradient descentin policy space, small-sample convergence results, algorithms and convergence results for partially observable MDPs, modular and hierarchical reinforcement learning.Recently, reinforcement learning has been used in the domain of Psychology to explain human learning and performance. In particular, it has been used in cognitive models that simulate human performance during problem solving and/or skill acquisition (e.g., Sun, Merril, & Peterson, 2001; Sun, Slusarz, & Terry, 2005; Gray, Sims, Fu, & Schoelles, 2006; Fu & Anderson, 2006). It was also used to propose a model of the human error-processing system (Holroyd & Coles, 2002). Multiagent or Distributed Reinforcement Learning is also a topic of interest in current research in this field.
Temporal difference learning
* cite journal
last = Kaelbling | first = Leslie P. | authorlink = Leslie P. Kaelbling
Michael L. Littman; Andrew W. Moore
title = Reinforcement Learning: A Survey
journal = Journal of Artificial Intelligence Research
volume = 4
pages = 237–285
date = 1996
url = http://www.cs.washington.edu/research/jair/abstracts/kaelbling96a.html
* cite book
last = Sutton | first = Richard S. | authorlink = Richard S. Sutton
Andrew G. Barto
title = Reinforcement Learning: An Introduction
publisher = MIT Press
date = 1998
isbn = 0-262-19398-1
url = http://www.cs.ualberta.ca/~sutton/book/ebook/the-book.html
id = refSutton1998
* cite book
last = Bertsekas | first = Dimitri P. | authorlink = Dimitri P. Bertsekas
title = Neuro-Dynamic Programming
publisher = Athena Scientific
date = 1996
location = Nashua, NH
isbn = 1-886529-10-8
url = http://www.athenasc.com/ndpbook.html
Ron Sun, E. Merrill, and T. Peterson, From implicit skills to explicit knowledge: A bottom-up model of skill learning. Cognitive Science, Vol.25, No.2, pp.203-244. 2001. http://www.cogsci.rpi.edu/~rsun/sun.cs01.pdf
Ron Sun, P. Slusarz, and C. Terry, The interaction of the explicit and the implicit in skill learning: A dual-process approach . Psychological Review, Vol.112, No.1, pp.159-192. 2005. http://www.cogsci.rpi.edu/~rsun/sun-pr2005-f.pdf
* cite conference
last = Peters | first = Jan
Sethu Vijayakumar; Stefan Schaal
title = Reinforcement Learning for Humanoid Robotics
booktitle = IEEE-RAS International Conference on Humanoid Robots
year = 2003
url = http://www-clmc.usc.edu/publications/p/peters-ICHR2003.pdf
* cite journal
last = Gray | first = Wayne D. | authorlink = Wayne D. Gray
Chris R. Sims; Wai-Tat Fu; Michael J. Schoelles
title = The Soft Constraints Hypothesis: A Rational Analysis Approach to Resource Allocation for Interactive Behavior
journal = Psychological Review
volume = 113
issue = 3
pages = 461–482
date = 2006
url = http://www.rpi.edu/~grayw/pubs/papers/GSFS06_PsycRvw/GSFS06_PsycRvw.htm
doi = 10.1037/0033-295X.113.3.461
format = Dead link|date=June 2008 – [http://scholar.google.co.uk/scholar?hl=en&lr=&q=author%3AGray+intitle%3AThe+Soft+Constraints+Hypothesis%3A+A+Rational+Analysis+Approach+to+Resource+Allocation+for+Interactive+Behavior&as_publication=Psychological+Review&as_ylo=&as_yhi=&btnG=Search Scholar search]
* cite journal
last = Fu | first = Wai-Tat | authorlink = Wai-Tat Fu
John R. Anderson
title = From Recurrent Choice to Skill Learning: A Reinforcement-Learning Model
journal = Journal of Experimental Psychology: General
volume = 135
issue = 2
pages = 184 –206
date = 2006
url = http://www.humanfactors.uiuc.edu/Reports&PapersPDFs/JournalPubs/Fu&Anderson06-JEPG(published)(ReinforcementLearning).pdf
doi = 10.1037/0096-34220.127.116.11
* [http://www-anw.cs.umass.edu/rlr/ Reinforcement Learning Repository]
* [http://rlai.cs.ualberta.ca/ Reinforcement Learning and Artificial Intelligence]
* [http://glue.rl-community.org RL-Glue]
* [http://www.dia.fi.upm.es/~jamartin/download.htm Software Tools for Reinforcement Learning (Matlab and Python)]
* [http://rlai.cs.ualberta.ca/RLR/index.html The UofA Reinforcement Learning Library (texts)]
* [http://www.igi.tugraz.at/ril-toolbox The Reinforcement Learning Toolbox from the (Graz University of Technology) ]
* [http://www.cogsci.rpi.edu/~rsun/hybrid-rl.html Hybrid reinforcement learning]
* [http://sourceforge.net/projects/piqle/ Piqle: a Generic Java Platform for Reinforcement Learning]
* [http://www.user.cifnet.com/~lwebzem/ttt/ttt.html Reinforcement Learning applied to Tic-Tac-Toe Game]
* [http://www.scholarpedia.org/article/Reinforcement_Learning Scholarpedia Reinforcement Learning]
Wikimedia Foundation. 2010.