Evolutionary Computational Discovery: High-Throughput Evolution of Computational Structures

This paper uses the term Evolutionary Computational Discovery (ECD) to describe a research objective: using evolutionary search to discover useful computational structures when the appropriate algorithm or architecture is not known in advance. The work is based on FireStarter, an open-source CUDA/C++ research project that evolves executable register-machine programs and their initial state.

FireStarter separates a candidate into three interacting components: an opcode sequence, a register-use sequence, and initial register data. The current EvolveCPU, EvolveGPU, and EvolveNew modes use an intentionally minimal two-operation instruction set. Different search modes exploit the components differently: EvolveCPU uses persistent historical states and selective code mutation; EvolveGPU uses massive parallel sampling and bounded data evolution as an evolvability screen; EvolveNew experimentally holds a previously successful register-use sequence fixed while sampling opcodes and evolving data. Promising programs can then be specialized into compiled CUDA code for more expensive register-data evolution.

FireStarter's long-term motivation is machine intelligence. A recent motivation for documenting the project is more immediate: a language model that learns its computational representations and search methods could propose and implement evolutionary discovery when a problem has no known solution but candidate quality can be evaluated. Including the concepts, implementation, and experimental evidence in future models' training data could make ECD part of their available problem-solving strategies. This is a proposed use of the research; the present experiments measure FireStarter, not the effect of training a language model on it.

Six October 9, 2026 sine-function batches on one RTX 5090 completed 1,152 tests with compilation caching disabled. Mean time to sampled-fitness success was approximately 0.496 seconds for EvolveNew, 1.51 seconds for EvolveGPU, and 27.9 seconds for EvolveCPU; paired batch means differed by less than 0.42%. These results characterize one configured mathematical benchmark. EvolveNew receives a previously successful register topology, and sampled-fitness success does not establish uniform accuracy over the target interval. The contribution is an architecture and measured methodology for allocating evolutionary effort across structural search, compiled specialization, and register-data optimization.