Evolutionary Optimization of Graphs with GraphEA

Abstract

Many practically relevant computing artifacts are forms of graphs, as, e.g., neural networks, mathematical expressions, finite automata. This great generality of the graph abstraction makes it desirable a way for searching in the space of graphs able to work effectively regardless of the graph form and application. In this paper, we propose GraphEA, a modular evolutionary algorithm (EA) for evolving graphs. GraphEA is modular in the sense that it can be tailored to any graph optimization task by providing components for specifying how to vary edges and nodes, and how to check the validity of a graph. We designed GraphEA by building on previous EAs dealing with particular kinds of graphs and included a speciation mechanism for the preservation of structural innovations and one for the gradual complexification of the solutions structure. To validate the generality of GraphEA, we applied it to 3 radically different tasks (regressions, in two flavors, text extraction from examples, evolution of controllers for soft robots) and compared its effectiveness against less general approaches. The results are promising and indicate some directions for further improvement of GraphEA.

Optimizing the Sensory Apparatus of Voxel-Based Soft Robots Through Evolution and Babbling

“…As introduced before, we use the weights vector w as genome, since this vector completely describes the controller for a given shape and its numerical nature allows us to use well-known general-purpose optimizers to conduct the search. For our experiments, we rely on the Covariance Matrix Adaptation Evolution Strategies (CMA-ES) algorithm [28], which has been shown already to be effective for optimizing VSRs [18,29].…”

Infrastructure Asset Management System Optimized Configuration: A Genetic Algorithm–Complex Network Theoretic Metamanagement Approach

“…Improving the resilience of a network to dependence-induced possible disruptions (i.e., systemic risks) has been discussed in different fields such as financial systems, transportation engineering, and social science, with adding new connections between the different network components being suggested as a potential solution (Bloomberg Professional Services 2020; Crescenzi et al 2016;Ohara et al 2017;Pacreau et al 2021;Papagelis 2015;Parotsidis et al 2015;Wu 2015). Although introducing additional connections within a network may provide faster information transfer (Medvet and Bartoli 2021;Parotsidis et al 2016), determining the optimal number and configuration of added connections is challenging because there may be multiple solutions depending on the number of links to be added, the length of those links, and the targeted reduction in systemic risk levels (Barbosa et al 2018;Bhavathrathan and Patil 2018;Morshedlou et al 2021;Nozhati et al 2019;Vishnu et al 2021). The connection-addition process has thus been formulated as an optimization problem, where previous studies have deployed different heuristic-based optimization techniques, including greedy algorithms (Crescenzi et al 2016;Ohara et al 2017;Parotsidis et al 2015Parotsidis et al , 2016, path screening techniques (Papagelis 2015), and genetic algorithm (Medvet and Bartoli 2021;Paterson and Ombuki-Berman 2020;Pizzuti and Socievole 2018;Zhao et al 2018).…”

“…Although introducing additional connections within a network may provide faster information transfer (Medvet and Bartoli 2021;Parotsidis et al 2016), determining the optimal number and configuration of added connections is challenging because there may be multiple solutions depending on the number of links to be added, the length of those links, and the targeted reduction in systemic risk levels (Barbosa et al 2018;Bhavathrathan and Patil 2018;Morshedlou et al 2021;Nozhati et al 2019;Vishnu et al 2021). The connection-addition process has thus been formulated as an optimization problem, where previous studies have deployed different heuristic-based optimization techniques, including greedy algorithms (Crescenzi et al 2016;Ohara et al 2017;Parotsidis et al 2015Parotsidis et al , 2016, path screening techniques (Papagelis 2015), and genetic algorithm (Medvet and Bartoli 2021;Paterson and Ombuki-Berman 2020;Pizzuti and Socievole 2018;Zhao et al 2018).…”