Minimalist Genetic Programming beats bloat with linguistic MERGE

Genetic programming (GP) traditionally treats program induction as an evolutionary search over syntax trees. In a new paper, Leonardo Trujillo flips this paradigm by proposing Minimalist Genetic Programming (MGP), which instead views the task as a syntactic derivation problem. Inspired by the Minimalist Program in linguistics, MGP replaces mutation and crossover with a single binary operation called MERGE, which combines atomic syntactic objects step-by-step using a simple Markovian process. This approach discovers core building blocks of symbolic expressions and constructs them incrementally without the uncontrolled growth (bloat) that often cripples standard GP.

Benchmarked on symbolic regression tasks known to be difficult for conventional GP, MGP consistently produces the exact ground truth model—something standard GP struggles to achieve. The results show that borrowing insights from human language syntax can offer a more disciplined, interpretable route to program induction. By framing the problem as a derivation rather than a search, MGP opens the door to more sample-efficient and bloat-free symbolic regression, with potential applications in scientific discovery, automated model building, and explainable AI.