The Effect of Laziness on Agents for Large Scale Global Optimization

Abstract

Practical optimization problems from engineering often suffer from non-convexity and rugged, multi-modal fitness or error landscapes and are thus hard to solve. Especially in the high-dimensional case, a lack of derivatives entails additional challenges to heuristics. High-dimensionality leads to an exponential increase in search space size and tightens the problem of premature convergence. Parallelization for acceleration often involves domain-specific knowledge for data domain partition or functional or algorithmic decomposition. On the other hand, fully decentralized agent-based procedures for global optimization based on coordinate descent and gossiping have no specific decomposition needs and can thus be applied to arbitrary optimization problems. Premature convergence can be mitigated by introducing laziness. We scrutinized the effectiveness of different levels of laziness on different types of optimization problems and for the first time applied the approach to a real-world optimization problem: to predictive scheduling in virtual power plant orchestration. The lazy agent approach turns out to be competitive and often superior to the non-lazy one and to standard heuristics in many cases including the real world problem.