Zoo Keeper RL – Reinforcement Learning Agent (PER-DQN & PPO)

A reinforcement learning project where an agent learns to manage a zoo — feeding animals, preventing escapes, and maximizing survival — built on a custom Gymnasium environment. Two RL algorithms were implemented, trained, and compared from scratch.

Key features:

Custom Gymnasium environment (ZooEnv) simulating a zoo grid with 6 animals, a 24-dimensional observation space (animal positions, hunger levels, agent position), and 5 discrete actions (movement + feeding)

Reward system designed around feeding hungry animals and penalizing escapes, with defined terminal conditions

Prioritized Experience Replay DQN (PER-DQN) implementation with a 100,000-sample replay buffer, epsilon-greedy exploration decay, and periodic target network updates, trained over 24,000 episodes

Proximal Policy Optimization (PPO) implementation using 8 parallel environments, Generalized Advantage Estimation (GAE), running normalization, and clipped policy updates, trained over 3,000,000 timesteps

Full training pipeline with CSV logging (rewards, escapes, losses, entropy/KL for PPO) and automatic GIF snapshot generation to visually track policy improvement over time

A visualization module (Vis.py) for plotting training metrics and comparing the two algorithms' performance