RL Agent

Mini Reinforcement Learning Library

Short Info

This is a mini reinforcement learning repo implementing simple agent for OpenAI gym.

This is work in progress.

Currently available agents:

AgentQ - e-greedy agent with q-function approximation

• Algorithm - not-configurable atm., currenty SARSA, other to be added later (e.g. Monte-Carlo, Eligibility Traces, Q-Learning)
• QFunction - pick from: QFunctTabular, QFunctAggregate, QFunctTiles
• Policy - must be PolicyEpsGreedy

AgentDQN - DQN agent, does everything as AgentQ, but includes reply memory as well

• Algorithm - not-configurable, always Q-Learning
• QFunction - should be non-linear QFunctKeras, specify keras model separately, other to be added later (?)
• Policy - must be PolicyEpsGreedy

AgentActorCritic - similar AgentQ, allows for different policies and state-value (V-Function) approximators

• Batching - not implemented - allow for collecting multiple episodes before learning
• Algorithm - non-configurable, currently Monte-Carlo, possibly add other later or even allow different alg. for actor and critic
• VFunction - not implemented, later add: VFunctTabular, VFunctAggregate, VFunctTiles
• QFunction - pick from: QFunctTabular, QFunctAggregate, QFunctTiles
• Policy - pick from: PolicyTabularCat, PolicyTabularCont, add Tiles and non-linear policies later

Future agents (possibly?)

AgentA2C - same as AgentActorCritic, but instead of batching it runs multiple agents in parallel on different copies of environment

Alternatives

If you are looking for more mature RL codebase, here are few choices:

• OpenAI baselines
• Deepmind Lab
• RLLab
• TensorForce
• keras-rl