thinkingML
Constructive Optimisation: From Sampling to Structure
Most hyperparameter optimisation methods rely on sampling fixed search spaces: defined ranges, heuristics, and distributions. Evolutionary Cellular Optimisation (ECO) departs from this norm entirely. It does not merely search - it builds.
ECO treats the hyperparameter landscape not as a pre-existing terrain to be explored, but as a generative object, one that is progressively shaped through interaction. At the core of this philosophy is the Holland von Neumann (HvN) landscape: a dynamic, structured representation where candidate solutions emerge from local rule-based evolution. These are not random guesses filtered post hoc, but constrained discoveries, configurations evolved through spatially local, temporally informed transformations.
This constructive view introduces significant advantages. First, the solution space adapts as the system learns, reflecting emergent structure rather than static assumptions. Second, the architecture enables ECO to work with extremely limited evaluation budgets. Since landscapes evolve incrementally, each new candidate is grounded in contextual information from prior performance, preserving continuity even under severe constraints.
Instead of drawing samples, ECO cultivates them. This distinction matters: in production environments where evaluations are costly and tolerances are tight, ECO can discover viable configurations faster, with greater robustness and transparency. Constructive optimisation is not just a new method, it's a new way of thinking about the problem itself.