Alpha2048 - how I taught my computer to play 1024
summary
Build trains and tests a 2048 solver using JavaScript libraries rather than theory it tries a supervised classifier with synaptic then switches to temporal difference Q learning with reinforced-js it experiments with rewards: raw score, a custom tile rating, and a score×rating heuristic it runs many simulated games overnight and shows modest score improvement but no consistent wins problems include flat 1D input losing board geometry, a single hidden layer in the library, and board crowding next steps are convolutional networks or porting to Python frameworks for deeper models
AI-written summary · about
chapters (12)
- 0:00Start
- 0:29LIAN LI
- 1:11THE GAME
- 2:52RULES
- 12:04REINFORCEMENT
- 15:08TEMPORAL DIFFERENCE
- 17:12REINFORCEJS
- 18:52FIRST ITERATION
- 19:44SECOND ITERATION
- 24:31THIRD ITERATION
- 26:29CONCLUSION
- 28:48NEXT
related talks
20mAlpha2048
29mReinforcement Learning in the Browser
43mAlpha 2048 - Reinforcement Learning in the Browser
49mML: Teaching machine to learn from its own mistakes (...)
18mJavaScript and Artificial Intelligence
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
29mHow I entered the machine learning world
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul