Hiding region constraints for black-box optimization. Application to camera placement in a virtual industrial environment using Evolutionary Computation

Abstract

Many real-world optimization problems related to physical environments have heavily constrained search spaces, which hinders the direct application of meta-heuristics and similar black-box methods. This work describes how to avoid region constraints and self-adapt search spaces without renouncing competitive solutions. The proposal relies on defining a gateway function that hides environment-specific placement constraints and is compatible with regular meta-heuristics and simulation-based optimization. The function can show a standard box-constrained domain encapsulating the real places involved. It has been successfully applied to automatic camera placement for task observation in a particle accelerator. The environment and the process of interest are simulated in the Unity game engine, which defines a cutting-edge trend in the design of such facilities. The primary optimization method tested is the genetic algorithm of MATLAB’s Global Optimization Toolbox, an industry standard that achieves remarkable results. The widespread Teaching–Learning-Based Optimizer (TLBO) and a random search have also been tried to complement the study. According to the results, the proposal does not prevent the advanced optimizers from finding camera arrangements that outperform (and are validated by) a human expert. It also allows the random search to find reasonable arrangements despite the underlying intricate set of constraints.