yaaf 0.0.1

YAAF: Yet Another Agents Framework

YAAF: Yet Another Agents Framework

A minimalistic reinforcement learning framework.

Installation

  $ pip install yaaf                               # Install YAAF
  $ pip install gym[atari]                         # Install OpenAI Gym's Atari environments
  $ git clone https://github.com/jmribeiro/yaaf    # Clone the repo for the examples and tutorials
  $ cd yaaf/examples                               
  $ python 0_space_invaders_random_agent.py        # Run example

• Rapid Prototyping: Setting up an agent on an environment can easily be done with a few lines of code:

  # Setup the Environment
  environment = OpenAIGymEnvironment(name="SpaceInvaders-v0", render=True)
  
  # Setup the Agent
  agent = RandomAgent(environment.action_space)
  
  # Run the agent on the environment for 5 episodes
  runner = EpisodeRunners(agent, environment, episodes=5)
  runner.run()
  environment.close()
  

• Simplicity

  States: Tensors/Numpy arrays with a given shape;

  Actions: Integers if a discrete action space, Floats if a continuous action space;

  Timesteps: Named tuples in shape of (state, action, reward, next_, is_terminal, info)

  Agents: Objects capable of execution actions when given a state. If trainable, capable of learning when given a timestep

  Environments: Objects which evolve over time, where agents can execute actions;

  Runners: Objects that interact an actor (agent or policy) on an environment, notifying a list of observers at every timestep;

  Metrics: Objects that evaluate an agent's or policy's performance on a given run. Passed as observables to the Runners.

  Presenters (TODO): Objects that display an agent's performance on a given run. Used to make plots and compare agents.

• Result Reproducibility:

  After creating and evaluating an agent on a given environment, it is possible to persistently save the agent to disk, allowing programmers to load up pre-trained agents and compare them in different scenarios.

  Implemented Agents:

  • DQN (https://storage.googleapis.com/deepmind-media/dqn/DQNNaturePaper.pdf)
  • QNetwork (http://ml.informatik.uni-freiburg.de/former/_media/publications/rieecml05.pdf)
  • QLearning (https://link.springer.com/content/pdf/10.1007%2FBF00992698.pdf)
  • SARSA

  Planned Agents:

  • [TODO] DDQN (https://arxiv.org/abs/1509.06461)
  • [TODO] A3C (https://arxiv.org/pdf/1602.01783.pdf)
  • [TODO] GA3C (https://arxiv.org/pdf/1611.06256.pdf)
  • [TODO] DDPG (https://arxiv.org/pdf/1509.02971.pdf)
  • [TODO] PPO (https://arxiv.org/abs/1707.06347)