Let's play Space Invaders!
summary
Reimplementation of Deep Q-Network (DQN) trains on raw grayscale frames (stacked 4×80×80) from the Arcade Learning Environment, learns actions {left,right,fire} by maximizing discounted rewards (gamma ≈ 0.99) with epsilon‑greedy exploration and experience replay, optimized with RMSprop. Training runs about 36–40 hours (~5 million actions) to reach ~1,000–1,500 points; best published variants hit ~9,000 with extra tricks. Code and emulator API are available and the agent behaves robustly though hyperparameters are sensitive.
AI-written summary · about
chapters (19)
- 0:00Reinforcement learning intro
- 1:04What is reinforcement learning
- 3:56Reinforcement learning in object detection
- 4:17Space Invaders with deep Q-learning
- 7:31Reward system in Q-learning
- 12:06Loss function for neural networks
- 14:18Multiple frames in gameplay
- 15:43Training with memory replay
- 16:47Exploration vs exploitation
- 18:14Challenges in Space Invaders
- 19:21Algorithm for other games like Doom
- 23:51AlphaGo vs Go masters
- 25:41Delayed rewards handling
- 27:33Hyperparameters gamma and epsilon
- 29:50Game environment setup
- 30:57Complex games with RL
- 32:06Real-world RL applications
- 33:13Advances in poker
- 35:04Q&A and conclusion
related talks
01h 01mReinforcement Learning: Pac-Man
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
01h 31mReinforcement Learning- Pac-Man
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
37mHow I trained an AI Model to Beat the 1990's Arcade Game Double Dragon
35mHow I Built an AI to Beat the 1990s Arcade Game Double Dragon
46mAI Takeover
31mUsing Python to build an AI to play and win SNES StreetFighter II