Usage Examples

In the following, we collect the links to MushroomRL scripts showing examples for most approaches available in MushroomRL.

The examples can be all found in the examples folder in the MushroomRL repository. They are grouped in four folders: papers reproduces experiments from a published paper, algorithms collects one script per algorithm, environments shows the interface of a specific environment class or simulator, and tools demonstrates the functionalities the library provides. Each of those folders carries its own README, describing its contents in more detail.

Every script can be launched directly, from any working directory:

python examples/algorithms/value/simple_chain_qlearning.py

Paper Reproductions

  • Grid World of Van Hasselt — Double Q-Learning, van Hasselt H., 2010

  • Double Chain — Q-Learning variants on the double chain of Relative Entropy Policy Search, Peters J. et al., 2010

  • Taxi with Mellowmax — An Alternative Softmax Operator for Reinforcement Learning, Asadi K. et al., 2017

  • CarOnHill with FQI — Tree-Based Batch Mode Reinforcement Learning, Ernst D. et al., 2005

  • CartPole with LSPI — Least-Squares Policy Iteration, Lagoudakis M. G. and Parr R., 2003

  • Atari with DQN — Human-Level Control Through Deep Reinforcement Learning, Mnih V. et al., 2015

These are the most expensive examples of the repository: several of them average over many independent runs, and the Atari one trains for tens of millions of frames at its published settings.

Value-Based Algorithms

Classical Policy Search and Actor-Critic

Black Box Optimization

Deep Actor-Critic

MuJoCo Environments

Continuous Control From Pixels

Vectorized Environments

The two multiprocess_environment scripts have a single-environment twin under algorithms/, so that diffing a script against its twin shows what vectorization costs.

Others Examples (Tools)