1.4.0

Changelog:

• Add full support for LOCM 1.5, including Gym envs for constructed phase (thanks, @lucca-nas) and the new battle phase rules.
• Add a consistency checker script to ensure gym-locm works exactly the same as the original engine.
• Major refactoring in the engine module.
• Convert draft training from stable-baselines to stable-baselines3.
• Remove full-game Gym envs.
• Remove tabular RL env, training script, and agent.
• Remove unfinished Coac and MCTS battle agents.
• Remove the tester script (the runner script has completely replaced it).
• Replace setup.py with pyproject.toml.
• Add the steps to reproduce the experiments from our newest Entertainment Computing paper.
• Use Black to reformat all code.
• Other bug fixes and minor changes.

Full Changelog: 1.3.0...1.4.0

1.3.0

Changelog:

• Add support to alternating between first and second players on the single-player and self-play battle envs.
• Add the role parameter to the training script: train your battle agent as the first player, second player, or alternating between first and second.
• Add the 'eval-battle-agents' parameter to the training script: evaluate your battle agent against any set of battle agents from agents.py.
• Add the steps to reproduce the experiments from our new SBGames and Entertainment Computing papers.
• Some fixes and quality of life changes in draft training.

Full Changelog: 1.2.0...1.3.0

1.2.0

Changelog:

• Add support for reward shaping
• Implement potential-based reward functions: win/loss, opponent health diff., player health diff., opponent board presence, player board presence, and Coac state evaluator
• Separate draft and battle training scripts (fixes #2)
• Add a hello world script for the battle envs
• Other bug fixes and minor changes

1.1.0

Changelog:

• Support different values of k and n in draft envs
• Add new draft models trained with 1M episodes
• Add the 'chad' draft agent (SCGAI competition 2020 winner)
• Add the 'historyless' draft agent (our best drafter so far)
• Support the use of stable-baseline3 algorithms on envs
• Use PPO with invalid action masking to train battle agents
• Log training metrics on Weights & Biases
• Many other minor changes and optimizations

1.0.0

Refactored and documented the envs. Contains reproduction-ready source code of our experiments with LOCM.

Thesis experiments version

This is the version used on the experiments on my thesis. It's fully functional, however a little refactoring and documenting would be ideal.