hjorvardr / rl_library

An implementation of main reinforcement learning algorithms: solo-agent and ensembled versions.

RL Library

Nowadays, artificial intelligence covers an important role in industry and scientific research. Next to clustering, deep learning and neural networks, reinforcement learning is becoming more and more popular. In the present work, the performance of reinforcement learning algorithms has been tested. Further more, two types of experiments have been performed:

• A solo-agent version, in which algorithms are executed as usual in the given environment.
• A cooperative version, in which two or more algorithms work together in order to take decisions.

Analysed algorithms

• Q-Learning
• SARSA
• DQN/DDQN
• AC (not fully tested)

Ensembling strategies

• Major voting based
• Rank voting based
• Trust based

OpenAI Gym

OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms written in Python. It provides a set of environments ranging from simple textual games to emulated Atari games and physics problems. Each environment is shipped with a set of possible actions/moves with a related reward. The user can use a standardised set of environments in order to feed the reinforcement learning algorithm. Moreover, an optional rendering is provided in order to offer a clear view of what is happening in background. There are different types of environments, characterised by different features such as:

• Observation space domain: discrete or continuous.
• Observation state type: memory representation or video frame.
• Reward range: finite or infinite set of values.
• Steps limitation.
• Maximum number of trials.

Testing environments

• Frozen-Lake4x4
• Frozen-Lake8x8
• Taxi
• MountainCar
• Breakout (not fully tested)
• Pong (not fully tested)
• CartPole