Pathfinding Neural Cellular Automata with Local Self-Attention

Abstract: Current artificial intelligence systems are rather rigid and narrow, if compared to the adaptivity and the open-endedness of living organisms. Neural Cellular Automata (NCA) are an extension of traditional CA, where the transition rule is replaced by a neural network operating on local neighborhoods. NCA provide a platform for investigating more biologically plausible features of emergent intelligence. However, an open question is how can collections of cells in an NCA be trained to collectively explore an environment in search for energy sources and find suitable paths to collect them. In this work, we utilize an NCA equipped with a local self-attention mechanism trained with gradient descent for pathfinding. Our results show that NCA can be trained to achieve such task and collect energy sources, while being able to redistribute the available energy to neighboring alive cells. Ongoing work is exploring how those abilities may be incorporated in NCA to solve tasks with increased adaptivity and general intelligence.