Abstract:
The field of evolutionary computation has drawn inspiration from Darwinian evolution in which species adapt to the environment through random variations and selection of the fittest. This type of evolutionary computation has found wide applications, but suffers from low efficiency. A recently proposed non-Darwinian form, called Learnable Evolution Model or LEM, applies a learning process to guide evolutionary processes. Instead of random mutations and recombinations, LEM performs hypothesis formation and instantiation. Experiments have shown that LEM may speed-up an evolution process by two or more orders of magnitude over Darwinian-type algorithms in terms of the number of births (or fitness evaluations). The price is a higher complexity of hypothesis formation and instantiation over mutation and recombination operators. LEM appears to be particularly advantageous in problem domains in which fitness evaluation is costly or time-consuming, such as evolutionary design, complex optimization problems, fluid dynamics, evolvable hardware, drug design, and others.
