---
tags:
- ml-intern
---
# TIL-26-AE Bomberman Agent — MaskablePPO + Curriculum Learning

This repository contains the training pipeline for an RL agent competing in the
**TIL-26 Automated Exploration** (AE) challenge — a competitive multi-agent
Bomberman-like environment.

## 🎯 Challenge

[Environment](https://huggingface.co/spaces/e-rong/til-26-ae): 2–6 team competitive Bomberman on a procedurally generated 16×16 maze. Key challenges:
- **Partial observability** (directional viewcones, not full map)
- **Sparse terminal rewards** (±50 for base destroy/survival)
- **Procedural generation** (new maze every episode)
- **Risk of camping** near base without exploration signal

## 🏗️ Architecture

### Three-Phase Training Pipeline

| Phase | Description | Opponents | Key Technique |
|---|---|---|---|
| **1** | MaskablePPO baseline | Random valid actions | Invalid action masking |
| **2** | Adaptive exploration | Random + visit-count bonus | Annealing: `α = 1 − tanh(k·deaths)` |
| **3** | Curriculum self-play | Rule-based (static → smart) | Elo-style difficulty progression |

### Design Decisions (Literature-Backed)

1. **MaskablePPO** (`sb3-contrib`): Handles invalid actions by setting logits to `-∞` before softmax. Proven superior to action penalties (Huang & Ontañón, 2020).
2. **MAPPO-style hyperparameters**: Value normalization, centralized value / decentralized policy, low sample reuse (Yu et al., NeurIPS 2022).
3. **Adaptive exploration annealing**: Directly from Pommerman SOTA (2024). As agent skill improves, exploration bonus decreases automatically, preventing camping.
4. **Curriculum learning**: 4 stages — static → simple → smart → mixed opponents. Advance at 55% win rate (or 500 episodes max).

### Key Papers

- **Pommerman multi-agent RL**: arxiv:2407.00662 — 98.85% win rate recipe
- **MAPPO best practices**: arxiv:2103.01955 — NeurIPS 2022
- **Invalid Action Masking**: arxiv:2006.14171 — theoretically justified
- **RND exploration** (fallback): arxiv:1810.12894 — if Phase 2 still camps

## 🚀 Running Training

### Prerequisites
```bash
# Download the environment (auto-bootstrapped in script)
python -c "from huggingface_hub import snapshot_download; snapshot_download('e-rong/til-26-ae', repo_type='space', local_dir='./til-26-ae-repo')"
```

### Local Training
```bash
export TOTAL_TIMESTEPS="500_000:500_000:1_000_000"
export HUB_MODEL_ID="E-Rong/til-26-ae-agent"
export TRACKIO_PROJECT="til-26-ae"
python train_all_phases.py
```

### HF Jobs (Recommended)
```bash
# Requires HF credits — run from a Space with the script uploaded
# Hardware: cpu-upgrade or a10g-large for GPU acceleration
```

## 📊 Monitoring

Trackio dashboard: `E-Rong/til-26-ae-trackio`

Logged metrics per phase:
- `train/mean_episode_reward`
- `train/mean_episode_length`
- `train/mean_explore_bonus` (Phase 2)
- `train/curriculum_stage` (Phase 3)

Alerts trigger on:
- Low reward (< -5) after 50k steps → suggests camping
- Curriculum stage advancement

## 📁 Repository Structure

```
train_all_phases.py   # Full 3-phase pipeline
requirements.txt      # Dependencies
bomberman_phase1_final.zip   # Saved after Phase 1
bomberman_phase2_final.zip   # Saved after Phase 2
bomberman_phase3_final.zip   # Saved after Phase 3
```

## 🧪 Evaluation

To evaluate a trained agent against random opponents:
```python
from train_all_phases import BombermanSingleAgentEnv
from sb3_contrib import MaskablePPO
from til_environment.config import default_config

cfg = default_config()
env = BombermanSingleAgentEnv(cfg=cfg)
model = MaskablePPO.load("bomberman_phase3_final")

obs, _ = env.reset(seed=42)
for _ in range(200):
    action, _ = model.predict(obs, action_masks=env.action_masks())
    obs, reward, done, truncated, info = env.step(action)
    if done or truncated:
        break
env.close()
```

## 📜 License

MIT — based on the TIL-26 AE challenge environment.

<!-- ml-intern-provenance -->
## Generated by ML Intern

This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub.

- Try ML Intern: https://smolagents-ml-intern.hf.space
- Source code: https://github.com/huggingface/ml-intern

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "E-Rong/til-26-ae-agent"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
```

For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class.