Final productive fitness is an a posteriori fitness estimate for evolutionary algorithms that takes into account the fitness of an individual's descendants. We use that metric in the context of surrogate-based evolutionary algorithms to show that computing a surrogate not for the original objective function but for the final productive fitness based on said objective function improves optimization properties of the surrogate function and might thus be a useful tool for certain dynamic optimization problems.
The natural-language-based genetic algorithm (NaLaGA) shows that a generic, problem-independent genetic algorithm (GA) can function by utilizing large language models (LLMs) for all interactions with the solution candidates without ever running any LLM-generated code. We analyze inherent biases in the function of a NaLaGA approach from prior work and propose some architectural changes to increase its generalizability.