Spatial Genetic Programming with the S1 Processing Board

This paper confirms the basic feasibility of implementing spatial genetic programming on the S1 Processing Board, a novel mesh-based massively parallel architecture optimized for high-throughput, energy-efficient computation. The S1 Board features broadcast-only instruction execution, fixed-point arithmetic, and a 2D mesh of 16 chips, each with over 2,000 processing elements for a total of approximately 33,000 elements. These unique hardwarehardwarecharacteristics require specialized algorithmic adaptations but enable the evolution of exceptionally large, spatially structured populations. We describe the design and deployment of an evolutionary algorithm on the S1, starting with a spatially distributed genetic algorithm applied to the MaxOnes problem, and extending to two more sophisticated genetic programming representations: Linear Genetic ProgrammingLinear Genetic Programming(LGP) and Cartesian Genetic ProgrammingCartesian Genetic Programming(CGP). Each approach is carefully adapted to exploit the S1’s architectural strengths, including synchronous population updates and local neighborhood selection through hardwarehardware-supported communication. Additionally, we present a modular software library that abstracts away the S1’s low-level features, simplifying the development of evolutionary algorithms on this platform. Our results confirm the S1’s capability to efficiently evolve large populations, highlighting its potential for advancing research in large-scale, spatially structured genetic programming.