On the Optimum Geometry and Training Strategy for Chemical Classifiers that Recognize the Shape of a Sphere

TL;DR

This paper explores how the geometry and training strategies of chemical oscillator networks can be optimized to improve their accuracy in classifying points inside a sphere, using evolutionary algorithms.

Contribution

It investigates the impact of network geometry and training data size on classifier performance, advancing the design of chemical oscillator-based classifiers.

Findings

Optimized network geometry can enhance classifier accuracy.

Training data size influences the effectiveness of evolutionary optimization.

Chemical oscillator networks can reliably classify points within a sphere.

Abstract

In this paper, we continue the discussion on database classifiers constructed with networks of interacting chemical oscillators. In our previous papers we demonstrated that a small, regular network of oscillators can predict if three random numbers in the range $[0,1]$ describe a point located inside a sphere inscribed within the unit cube $[0,1] \times [0,1] \times [0,1]$ with the accuracy exceeding $80 \%$. The parameters of the network were determined using evolutionary optimization. Here we apply the same technique to investigate if the classifier accuracy for this problem can be improved by selecting a specific geometry of interacting oscillators. We also address questions on the optimum size of the training database for evolutionary optimization and on the minimum size of the testing dataset for objective evaluation of classifier accuracy.