Reinforcement Learning and Minecraft
summary
Demonstrates using project Malmo (open source, no Minecraft purchase required) to train agents on a cliff task with rewards: move -1, blue +100, lava -100. Shows three stages: random actions, reward memory, and Q-learning (alpha=1.0, gamma=0.8) where the Q-table converges to a safe path; changing to alpha=0.5/gamma=0.9 slows convergence. Moves to deep RL using image-difference inputs and a convolutional policy (CNN: 16@8x8 stride4, 32@4x4 stride2, 256 FC, 12 actions), which succeeds on some seeds but overfits to specific layouts.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:52Minecraft
- 2:18What is Minecraft
- 3:10Creative and building aspects
- 4:32Project Mountain
- 7:36Supervised
- 7:58Unsupervised
- 8:42Agent
- 9:49Trial and Error
- 10:48Reinforcement Learning
- 11:44Falling to Lava
- 18:39Matrix
- 19:10Formula
- 21:02Example
- 26:59Deep Reinforcement Learning
- 29:23Subtracting Images
- 30:22Classification
- 31:52Code
- 32:39Demo
- 36:53Takeaway
- 39:53Book
- 40:36GitHub
related talks
32mAI and Minecraft
28mAI and Minecraft Lars
01h 09mMinecraft as AI Playground and Laboratory
50mIEEE CoG 2019 - Industry Day - Katja Hofmann (Microsoft Research)
05mInterview with
49mA Raspberry Pi and DeepLearning4J Brain That Plays Minecraft
50mA Raspberry Pi and DeepLearning4J Brain That Plays Minecraft
26mChatGPT and Minecraft's diamonds playing agents: Reinforcement Learning in a nutshell with sheeprl