learnrl 0.1.0

A package to learn about Reinforcement Learning

LearnRL

A library to use and learn reinforcement learning algorithms

General framework for reinforcement learning agents

Every agents can be implemented as a subclass of this general agent ! QLearning, DeepQLearning, A3C based agents ... and this librairie tries to prove it !

Memory

All the agents share the same memory system optimized with numpy to speedup computation See agents.agent for more details

Table Agent (With any controls or evaluations) including the famous QLearning !

Table agents builds a table of actions values Q(state, action) From this table the agent uses a Control algorithm to decide what action to play in each state: Control: state(s) -> action to play

Throught experience it updates this table using an evaluation algorithm: Evaluation: experience -> Q update

The QLearningAgent is already pre-built and ready for use ! Just import it from agents !

See agents.basic.agent for more details

Control

Commons controls are built-in : Greedy, UpperConfidenceBound(UCB), PolynomialUpperConfidenceTrees(Puct) But you can build any control algorithm by extanding the class Control ! You just have to code the function policy: policy(state(s), action_values(Q), action_visits(N)) -> probability of each action

See agents.basic.control for more details

Evaluation

Commons evaluations are built-in : MonteCarlo, TemporalDifference(lambda) But you can build any control algorithm by extanding the class Evaluation ! You just have to code the function learn: learn(action_values(Q), action_visits(N), memory:Memory) -> update Q and N

See agents.basic.evaluation for more details