go back

Inductive Biases and Metaknowledge Representations for Search-based Optimization

Stephen Friess, "Inductive Biases and Metaknowledge Representations for Search-based Optimization", University of Birmingham, 2022.


"What I do not understand, I can still create.", H. Sayama. The following work follows closely the aforementioned bonmot. Guided by questions such as: "How can evolutionary processes exhibit learning behavior and consolidate knowledge?", "What are cognitive models of problem-solving?" and "How can we harness these altogether as computational techniques?", we clarify within this work essentials required to implement them for metaheuristic search and optimization. We therefore look into existing models of computational problem-solvers and compare these with existing methodology in literature. Particularly, we find that the meta-learning model, which frames problem-solving in terms of domain-specific inductive biases and the arbitration thereof through means of high-level abstractions resolves outstanding issues with methodology proposed within the literature. Noteworthy, it can be also related to ongoing research on algorithm selection and configuration frameworks. We therefore look in what it means to implement such a model by first identifying inductive biases in terms of algorithm components and modeling these with density estimation techniques. And secondly, propose methodology to process metadata generated by optimization algorithms in an automated manner through means of deep pattern recognition architectures for spatio-temporal feature extraction. At last we look into an exemplary shape optimization problem which allows us to gain insight into what it means to apply our methodology to application scenarios. We end our work with a discussion on future possible directions to explore and discuss the limitations of such frameworks for system deployment.

Download Bibtex file Per Mail Request