MiniGrid-RL
Idokorro/MiniGrid-RL ↗My PhD codebase. It trains agents on several task families in the MiniGrid gridworld and compares ways to get a single agent that handles all of them:
- Single-task and multi-task PPO baselines, plus fine-tuning a multi-task model on one task.
- Policy distillation from task-specific experts into one student policy.
- Mixture of Experts, which routes between specialist policies. It scored best on the combined benchmark (72% vs 65% for multi-task PPO).
- LLM-driven agents running local models through Ollama, compared against the learned policies.
- Hyperparameter search with custom Hydra sweepers for PBT, PB2 and DEHB.