SPURL is an open-source toolkit for building self-play algorithms to solve reinforcement learning environments. SPURL's modular build allows users to train with SPURL for a variety of self-play, multi-agent or single-agent reinforcement learning problems.
We present a variety of demos to illustrate the functionality of SPURL:
To be added and/ or tested: Connect Four Pong Pendulum Bipedal Walker Soccer Twos
Added/ tested demos:
Environment Type | Example (Solved by SPURL) |
---|---|
Single-Agent Discrete Actions (Cartpole) | |
Self-Play Discrete Actions (TicTacToe) |
SPURL currently demonstrates the following functionality:
Feature | Support |
---|---|
Action Space | Discrete/ Continuous |
Opponent Sampling | Vanilla/ Ficticious/ Prioritised Ficticious |
RL Scenarios | Self-Play/ Co-op MARL/ Single-Agent |
Opponent Experience for Training | Yes/ No |
Entry-points into SPURL are present in spurl.core
and may be used to train or test any algorithms.
Please see demos
for example usage to train or test using self-play environments.