Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN
Researchers have partially reverse-engineered a convolutional recurrent neural network (RNN) used for the game Sokoban. They discovered that the network stores future moves, or plans, as activations within specific "path channels" in its hidden state. These channels are influenced by convolutional kernels that encode learned transition models, allowing the RNN to construct plans by propagating activations from boxes to goals and using negative values to prune paths at obstacles, effectively implementing a form of backtracking. AI