Incremental Evaluation in Genetic Programming.

Often GP evolves side effect free trees. These pure functional expressions can be evaluated in any order. In particular they can be interpreted from the genetic modification point outwards. Incremental evaluation exploits the fact that: in highly evolved children the semantic difference between child and parent falls with distance from the syntactic disruption (e.g. crossover point) and can reach zero before the whole child has been interpreted. If so, its fitness is identical to its parent (mum). Considerable savings in bloated binary tree GP runs are given by exploiting population convergence with existing GPquick data structures, leading to near linear O(gens) runtime. With multi-threading and SIMD AVX parallel computing a 16 core desktop can deliver the equivalent of 571 billion GP operations per second, 571 giga GPop/s. GP convergence is viewed via information theory as evolving a smooth landscape and software plasticity. Which gives rise to functional resilience to source code changes. On average a mixture of 100 +, -, x and (protected) divided by tree nodes remove test case effectiveness at exposing changes and so fail to propagate crossover infected errors.