ROCCYK/RL_Space_Invaders
An interactive reinforcement learning project where agents trained using Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) play Atari's classic Space Invaders. The project includes a Streamlit app to visualize and compare the performance of both agents.
What's novel
An interactive reinforcement learning project where agents trained using Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) play Atari's classic Space Invaders. The project includes a Streamlit app to visualize and compare the performance of both agents.
Code Analysis
7 files read · 3 roundsImplements textbook DQN and PPO agents from scratch in PyTorch to play Atari Space Invaders, with a Streamlit app to visualize gameplay and benchmark 100-game average rewards between the two agents.
Strengths
Both RL algorithms are implemented correctly from scratch with proper components (experience replay, target networks, GAE, clipped surrogate objective), and the training logs confirm the agents actually learned to play. The Streamlit app provides a clean, functional interface for real-time visualization and head-to-head comparison.
Weaknesses
No tests exist, the DQN epsilon decay is misconfigured (1M steps means the agent is still 73% exploratory at training end), the PPO rollout buffer has a subtle state-reset bug when episodes terminate mid-rollout, and there's significant code duplication (FrameStack, preprocess, and network architecture are copy-pasted between DQN.py and PPO.py).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
12
Contributors
1
Files
17
Active weeks
2
Repository
Language
Python
Stars
1
Forks
0
License
—