Training Reinforcement Learning Agents on Street Fighter III Using a MAME Wrapper Python Library

The article introduces a Python library that serves as a wrapper for the MAME emulator, enabling researchers and developers to train reinforcement‑learning algorithms on classic arcade games, with Street Fighter III used as a concrete example.

System requirements : a Linux operating system and Python version 3.6 or newer.

Installation : the library can be installed via pip install <library-name> (the exact package name is shown in the accompanying image).

Basic usage example : after installation, a short script demonstrates how to create a random agent that interacts with the Street Fighter III environment, showing how the library provides frame data, game state, and memory addresses for each step.

Debugging and game identification : the MAME debugger can be launched to obtain the game ID (visible in the title’s parentheses) and to inspect memory addresses. Commands for opening the debugger and retrieving the ID are described, and screenshots illustrate the process.

Game environment configuration : users can list available input actions, adjust the frame_ratio to control the simulated frame rate (default 60 fps, adjustable to 20 fps with a simple code change), and specify which memory addresses to monitor.

Step function and interaction : the library’s step function returns the current frame and associated data as NumPy matrices, along with all memory address values for the current timestep. It also allows sending actions to the emulator by specifying the appropriate input ports and fields.

Performance benchmarking and a simple ConvNet agent : the article presents benchmark results for the simulation speed and provides a minimal convolutional neural network agent that learns to play the game, with visual results shown in the included images.

Disclaimer : the content is compiled from online sources, and copyright belongs to the original authors; any errors or infringements should be reported for removal or permission.