Alpha2048
summary
Built a browser prototype called Alpha2048 that uses JavaScript libraries synaptic and reinforced.js with Q‑learning. Data collection starts from serialized 4×4 states; reward engineering (normalized score × normalized moves) and decaying epsilon improve outcomes. Adding a second hidden layer yields slow but visible learning over hours or days; trained agent outperforms random play yet stalls on crowded 4×4 boards because inputs are flattened and spatial structure is lost.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:35LIAN LI
- 2:08THE GAME 2048
- 2:27RULES
- 6:25GATHER DATA
- 9:41REINFORCEMENT LERON
- 10:31TEMPORAL DIFFERENCE
- 11:57REINFORCEJS
- 13:59FIRST ITERATION
- 15:07SECOND ITERATION
- 16:19THIRD ITERATION
related talks
29mReinforcement Learning in the Browser
33mAlpha2048 - how I taught my computer to play 1024
43mAlpha 2048 - Reinforcement Learning in the Browser
26mGames Are Smarter Than Us
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
18mJavaScript and Artificial Intelligence
56mCan I do AI?
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul