LEGA: an architecture for learning and evolving flexible job-shop schedules

The interaction between evolution and learning has received much attention with recent studies in machine learning showing that it can significantly improve the efficiency of evolutionary strategies for job-shop scheduling. We propose a tripartite architecture called LEGA; comprising a population generator that improves the quality of the initial population for subsequent evolution while training a schemata learning module to modify the fitnesses of its offsprings aided by a memory of conserved schemas resolved from sampled schedules received dynamically during evolution. Experimental results indicate that an instantiation of LEGA outperforms current approaches using canonical EAs in computational time and quality of schedules.