This project studies decision-making under uncertainty in aquatic environments using single-agent and cooperative multi-agent reinforcement learning (MARL).
The environment simulates stochastic waste generation, flow-driven transport, and interception by mobile agents. The goal is to maximize total intercepted waste over an episode.
- Custom aquatic environment with stochastic dynamics
- Random and greedy heuristic baselines
- Single-agent PPO (Stable-Baselines3)
- Cooperative multi-agent PPO (centralized training, joint actions)
- Quantitative comparison of policies
- Grid-based aquatic system
- Upstream waste spawning
- Downstream flow and drift
- Local interception radius
Each episode represents an independent stochastic simulation.
Mean episodic reward (± std):
- Random (single-agent): ~12.7
- PPO (single-agent): ~19.6
- PPO (2-agent cooperative): ~27.4
- Greedy heuristic: ~30.1
Cooperative MARL significantly improves performance over single-agent learning, approaching a strong heuristic baseline.
pip install -r requirements.txt