Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN

A convolutional recurrent neural network (RNN) has been partially reverse-engineered to understand its planning mechanisms while playing the box-pushing game Sokoban. The study reveals that the RNN utilizes specific channels, termed path channels, to store future moves, with activations indicating the direction boxes will be pushed based on their locations.

This development provides insights into how RNNs can encode planning strategies and transition models, potentially enhancing the understanding of artificial intelligence in complex task environments and improving future AI applications in similar domains.