Aff54/Reinforcement-learning-applied-to-Liar-s-dice
Trying to train agents to play Liar's dice with DDQN and masking.
What's novel
Trying to train agents to play Liar's dice with DDQN and masking.
Code Analysis
9 files read · 5 roundsImplements a Liar's Dice game environment with rule-based opponents, probability-based decision-making, and RL infrastructure (replay buffer, epsilon-greedy agent, action masking), but the actual DDQN network and training loop are only in a notebook, not in the source package.
Strengths
The game engine is well-structured with clean separation between base Game and GameRL classes, the action masking logic correctly implements all Liar's Dice outbidding rules including wild/non-wild transitions, and the probability analysis with proper binomial calculations is mathematically sound and well-tested.
Weaknesses
The core claimed contribution (DDQN implementation with target network and training loop) is absent from the source package — only the inference agent and replay buffer exist in src/, making the 'DDQN implementation' claim in the README misleading; the @lru_cache on functions with mutable list arguments in deterministic_agents.py is a subtle bug that will cause incorrect caching behavior.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
115
Contributors
1
Files
19
Active weeks
12
Repository
Language
Python
Stars
3
Forks
0
License
MIT