I competed together with Kota Iizuka. This repository only contains the code of my approach.

• Approach
  • BestGuess
  • Optimizer
• Notebook

Given a game of life end state, the task is to find a valid start state, which evolves to the desired end state after delta steps.

The approach has two main components:

• BestGuess, a neural network which predicts the start state for a single step. It is repeated for delta greater than 1.
• Followed by an Optimizer

The neural network only predicts one cell at a time, as there are many possible start state which lead to the same end state. Inspired by language models, it gets the already predicted cells as input to create an in itself consistent result.

The optimizer picks a random 3x3 region in the predicted start state, tries every possible combination of the 3x3 cells at once and selects randomly between the ones which produce the best score. It is entirely implemented on the GPU.

An animation of the approach solving a state can be seen in the notebook on kaggle.