Reinforcement Learning

From Chessprogramming wiki
Revision as of 12:21, 25 August 2018 by GerdIsenberg (talk | contribs)
Jump to: navigation, search

Home * Learning * Reinforcement Learning

Reinforcement Learning,
a learning paradigm inspired by behaviourist psychology and classical conditioning - learning by trial and error, interacting with an environment to map situations to actions in such a way that some notion of cumulative reward is maximized. In computer games, reinforcement learning deals with adjusting feature weights based on results or their subsequent predictions during self play.

Reinforcement learning is indebted to the idea of Markov decision processes (MDPs) in the field of optimal control utilizing dynamic programming techniques. The crucial exploitation and exploration tradeoff in multi-armed bandit problems as also considered in UCT of Monte-Carlo Tree Search - between "exploitation" of the machine that has the highest expected payoff and "exploration" to get more information about the expected payoffs of the other machines - is also faced in reinforcement learning.

Q-Learning

Q-Learning, introduced by Chris Watkins in 1989, is a simple way for agents to learn how to act optimally in controlled Markovian domains [2]. It amounts to an incremental method for dynamic programming which imposes limited computational demands. It works by successively improving its evaluations of the quality of particular actions at particular states. Q-learning converges to the optimum action-values with probability 1 so long as all actions are repeatedly sampled in all states and the action-values are represented discretely [3]. Q-learning has been successfully applied to deep learning by a Google DeepMind team in playing some Atari 2600 games as published in Nature, 2015, dubbed deep reinforcement learning or deep Q-networks [4], soon followed by the spectacular AlphaGo and AlphaZero breakthroughs.

Temporal Difference Learning

see main page Temporal Difference Learning

Q-learning at its simplest uses tables to store data. This very quickly loses viability with increasing sizes of state/action space of the system it is monitoring/controlling. One solution to this problem is to use an (adapted) artificial neural network as a function approximator, as demonstrated by Gerald Tesauro in his Backgammon playing temporal difference learning research [5] [6].

Temporal Difference Learning is a prediction method primarily used for reinforcement learning. In the domain of computer games and computer chess, TD learning is applied through self play, subsequently predicting the probability of winning a game during the sequence of moves from the initial position until the end, to adjust weights for a more reliable prediction.

See also

UCT

Selected Publications

1954 ...

1960 ...

1970 ...

1980 ...

1990 ...

1995 ...

2000 ...

2005 ...

2010 ...

2011

2012

István Szita (2012). Reinforcement Learning in Games. Chapter 17

2013

2014

2015 ...

2016

2017

Postings

External Links

Reinforcement Learning

MDP

Q-Learning

Courses

  1. Lecture 1: Introduction to Reinforcement Learning
  2. Lecture 2: Markov Decision Process
  3. Lecture 3: Planning by Dynamic Programming
  4. Lecture 4: Model-Free Prediction
  5. Lecture 5: Model Free Control
  6. Lecture 6: Value Function Approximation
  7. Lecture 7: Policy Gradient Methods
  8. Lecture 8: Integrating Learning and Planning
  9. Lecture 9: Exploration and Exploitation
  10. Lecture 10: Classic Games

References

Up one Level