You may have heard of Heroes of Might and Magic III, a classic strategy and tactics game featuring a rule-based AI that is relatively weak.
The vcmi-gym project allows training of machine-learning models, enabling the first-of-its-kind HOMM3 AI powered by reinforcement learning.
vcmi-gym offers a gym-compatible reinforcement learning (RL) environment specifically designed for VCMI, the open-source engine recreating the classic game Heroes of Might and Magic III.
It includes implementations of various RL algorithms and supplementary tools that support model orchestration, hyperparameter tuning, and observability, enabling the creation of AI combat models for VCMI.
Trained models from vcmi-gym are featured in VCMI as of December 2025. Players can now choose "MMAI" in VCMI's settings to enable the ML-powered AI which uses vcmi-gym models, enriching gameplay experience during battles.
Feel free to try it out! Install VCMI, choose "MMAI" and press Play ;)
vcmi-gym is in a state of active development. Updates to the
gym environment, NN architecture and RL algorithms are made frequently.
Although the first models are already out, training new models is an
ongoing effort.
Read more about this project here.
No comments yet.
Sign in to be the first to comment.