Self-adaptive (mu, lambda) Evolution Strategy over a real-valued vector.

The first optimizer of Programme 2 (Search Strategies). Unlike the DXNN path's truncation-selection GA + memetic tuner, this is a canonical ES: each individual is a parameter vector plus its own mutation step size sigma, sigma self-adapts log-normally, and selection is COMMA (the mu best of the lambda offspring become the next parents; parents do not survive). This is the method family that solves the pole-balancing benchmarks the GA stalled on (insight 024; Igel 2003, Gomez CoSyNE 2008).

The fitness function is pluggable: evolve/3 takes fun(([float()]) -> float()), higher is better. That lets the ES be validated on a benchmark function (mu_lambda_es_tests) independently of network evaluation; wiring it to a fixed-topology network's weight vector is EXP-025.

Reference: Beyer and Schwefel 2002, "Evolution strategies - a comprehensive introduction"; Rechenberg / Schwefel.