Curious-Transformer-On-Rubik-Cube

First time setup

Install the hydra-core and transformers packages (ontop of the usual pytorch and numpy)

File structure:

The important folders are:

• CubeGPT: The Regular Transformer
• Decision Transformer: The decision transformer and online decision transformer Please enter these directories for more detailed README instructions to run the code.

Name

• Curious Transformer.
  • Curious: Curiousity-driven learning.
  • Transformer: Attention mechanism.

Core Idea:

• Transformer $\in$ Reinforcement Learning $\rightarrow$ Decision Transformer.
• Decision Transformer (offline) + Curiousity-driven Learning $\rightarrow$ Curious Transformer (online).
• No expert data.
• No problem-specific, hard-coded solver.
• No overkill with models like GPT-3.
• Apply Curious Transformer to solve $3 \times 3$ Rubik's cube.
• Generalize to $d \times d$ Rubik's cube.
• Generalize to any Reinforcement Learning tasks with deterministic, fully observable, discrete environment.