So, this is the project to train a Superhuman bot for Disciples 2, popular turn-based strategy game. There are 3 stages:

1. Single-player scenarios First, an agent is trained to complete individual single-player scenarios. Autoresearch is used to test multiple configurations. Preliminary roadmap

2. Randomly generated and unseen maps Next, a more general agent is trained to play previously unseen maps, face randomly generated enemies, and handle a wider variety of situations.

3. Self-play training Finally, an AlphaZero-like agent is trained through self-play, with the goal of competing against the best Disciples II players and eventually beating the champion—whoever that may be. Why not just use AlphaZero right away?

All features of the original game have been implemented in Python, except for rod planters, the recruitment of new heroes, the hiring of thieves, non-interactive map objects, сastle guard and the original visuals.

You can run agent training for a real Disciples 2 scenario "A Return To Simpler Times" (with the exceptions above) with the following command. By default, a slightly modified Maskable PPO algorithm from Stable Baselines 3 Contrib is used.

python Big_map/train_campaign.py

The agent is being trained for Disciples II: Rise of the Elves: Disciples II: Rise of the Elves on Steam