Reinforcement Learning in the Browser
summary
Prototype Alpha2048 runs entirely in the browser, starts with synaptic then moves to reinforceJS, and plays the 4x4 2048 game (new tiles 90% 2, 10% 4). Training experiments compare using raw score versus delayed reward and an epsilon decay exploration schedule. The agent shows modest learning but quickly plateaus; implementation notes include a manually added second layer, heavy CPU in-browser training, and an opaque reward design that limits foresight.
AI-written summary · about
chapters (15)
- 0:00Start
- 4:22Rules
- 7:13Constraints
- 7:46Difference between Supervised Learning and Unsupervised Learning
- 8:24Unsupervised Learning
- 9:49Gather Data
- 10:02Simulate all Possible Actions
- 11:17Inputs and Outputs
- 13:05Conclusion
- 15:20Delayed Reward
- 15:47Exploration versus Exploitation
- 16:56Temporal Difference Learning
- 21:40Decaying Epsilon over Time
- 25:41Cap on the Board Size
- 27:29Demo
