A Program to digitize many twisty puzzles and solve them with an AI using Reinforcement Learning and automatic algorithm generation.
This project includes a new twisty puzzle simulator for permutation puzzles as well as different solvers. The program can detect rotational symmetries of puzzles, find useful algorithms to solve them and then train an AI to solve the puzzle using Reinforcement Learning, making use of the generated algorithms and rotational symmetries.
Development started using python 3.8.5, last tested in 3.11.3
third party modules:
see requirements.txt for a list of all required modules and versions.
All of them can be easily installed with pip install ... or pip install -r requirements.txt to install all of them at once. You may need to install NVIDIA CUDA toolkit beforehand. Otherwise the program will run on the CPU which is slower.
The repository contains a folder doc which includes markdown files with explanation of the implementation of many features.
Currently the program is supposed to be used via a command-line interface. A proper user interface is planned but can still take a long time. There is no plan when I will find time to implement it.
This version focusses on the 3d animation using vpython
It also implements a console interface to control the animation. A list of all CLI commands can be accessed by running run_program.py and typing help (then pressing enter).
See doc/user guide.md for more detailed usage instructions.
- DONE: Much better AI that can find and utilize useful algorithms for any puzzle and solve more complex puzzles.
- Simplify puzzle definition. -> Partially done. Cuboid generation is now automated with
src/puzzles/cuboid_generator.py - A UI with a 3D viewport as well as buttons to control the shown puzzles and the AI solving them.
- Enable user to save Algorithms for puzzles and use them to solve the puzzle more easily.
- Save written solving strategies for puzzles.
- Enable user to save the current state of a puzzle and load it again later.