Surrogate-Assisted Genetic Programming with Phenotypic Characterisation in Dynamic Multi-Mode Project Scheduling

Details

Genetic programming (GP) can evolve scheduling heuristics for dynamic multi-mode resource-constrained project scheduling, but simulation-based fitness evaluation is computationally expensive. Surrogate models offer cheaper fitness estimates, but require careful design of how GP individuals are encoded into phenotypic characterizations for the surrogate to be useful. This paper tests three encoding schemes: priority-value (raw rule outputs), rank (candidate ordering), and binary (final scheduling decisions), combined with three distance metrics. Binary encoding with Euclidean distance consistently outperforms alternatives on surrogate guidance effectiveness. A key finding is that surrogate estimation accuracy alone does not explain performance differences — the representation also controls how effectively phenotypically redundant offspring are pruned and how behavioral diversity is retained after preselection. Ablation experiments confirm that duplicate removal and surrogate preselection provide complementary benefits, with their combination yielding the largest improvement. This is niche evolutionary computation research with applications to combinatorial optimization in project management, manufacturing, and logistics. No implementation is released and no enterprise-facing tool is described.