Grammar‐based autonomous discovery of abstractions for evolution of complex multi‐agent behaviours

Abstract

This paper presents a grammar-based evolutionary approach that facilitates autonomous discovery of abstractions to learn complex collective behaviours through manageable sub-models. We propose modifications to the design of the genome structure of the evolutionary model and the grammar syntax to facilitate representation of abstractions in separate partitions of a genome. Two learning architectures based on parallel and incremental learning are proposed to automatically derive abstractions. The evaluations conducted with three different complex task environments indicate that the proposed approach with both architectures surpass the performance of generic grammar-based evolutionary models by automatically identifying appropriate abstractions and generating more complex rule structures. The evolutionary process shows further performance improvements with the use of scaffolded environments which were used to train the models in increasingly complex environments across several stages. The results infer that the proposed approach incorporating grammatical evolution with techniques to autonomously discover abstractions can facilitate solving complex problems of agent systems in real-world domains.