Reversing Conway’s Game of Life Jonathan Goetz
The Rules of the Game 1 living neighbor = cell dies 2 living neighbors = cell maintains 3 living neighbors = cell lives 4-8 living neighbors = cell dies Many states share descendants Some states have no ancestors Some states are their own ancestors Information is lost In Reverse?
Challenges Loss of information prevents traditional back calculation of initial state The problem itself maps rather easily to a 400 input and 400 output MLP pattern classification which is applied recursively.
Details The basic structure is a multilayered set of MLPs which rescales the result to a range of 0- 1 before reapplying the same MLP in order to approximate the input layout. Initial attempts to create a variable depth training weighting were not successful.
Status Initial attempts at rewriting the training process for between 1-5 levels of difference unsuccessful. Instead implementing this by converting every multi-generational training data point into multiple 1 generation steps.