FelipeMarcelino/2048-Gym

Implements a custom OpenAI Gym environment with DQN training via Stable Baselines, supporting dual board representations (binary and raw) and feature extraction through either CNN or MLP architectures. Training is accelerated using Numba's JIT compilation, while hyperparameter optimization is automated through Optuna's define-by-run framework. The CNN-based agent achieved the 2048 tile in 10% of test games, demonstrating superior spatial feature learning compared to MLP approaches.

No commits in the last 6 months.