This study presents a new evolutionary network minimization (ENM) algorithm. Neurocontroller minimization is beneficial for finding small parsimonious networks that permit a better understanding of their workings. The ENM algorithm is specifically geared to an evolutionary agents setup, as it does not require any explicit supervised training error, and is very easily incorporated in current evolutionary algorithms. ENM is based on a standard genetic algorithm with an additional step during reproduction in which synaptic connections are irreversibly eliminated. It receives as input a successfully evolved neurocontroller and aims to output a pruned neurocontroller, while maintaining the original fitness level. The small neurocontrollers produced by ENM provide upper bounds on the neurocontroller size needed to perform a given task successfully, and can provide for more effcient hardware implementations.

One of the major challenges in the field of neurally driven evolved autonomous agents is deciphering the neural mechanisms underlying their behavior. Aiming at this goal, we have developed the multi-perturbation Shapley value analysis (MSA)—the first axiomatic and rigorous method for deducing causal function localization from multiple-perturbation data, substantially improving on earlier approaches. Based on fundamental concepts from game theory, the MSA provides a formal way of defining and quantifying the contributions of network elements, as well as the functional interactions between them. The previously presented versions of the MSA require full knowledge (or at least an approximation) of the network's performance under all possible multiple perturbations, limiting their applicability to systems with a small number of elements. This article focuses on presenting new scalable MSA variants, allowing for the analysis of large complex networks in an efficient manner, including large-scale neurocontrollers. The successful operation of the MSA along with the new variants is demonstrated in the analysis of several neurocontrollers solving a food foraging task, consisting of up to 100 neural elements.

This article investigates the evolution of autonomous agents that perform a memory-dependent counting task. Two types of neurocontrollers are evolved: networks of McCulloch-Pitts neurons, and spiking integrate-and-fire networks. The results demonstrate the superiority of the spiky model in evolutionary success and network simplicity. The combination of spiking dynamics with incremental evolution leads to the successful evolution of agents counting over very long periods. Analysis of the evolved networks unravels the counting mechanism and demonstrates how the spiking dynamics are utilized. Using new measures of spikiness we find that even in agents with spiking dynamics, these are usually truly utilized only when they are really needed, that is, in the evolved subnetwork responsible for counting.

This article presents a novel method for the evolution of artificial autonomous agents with small neurocontrollers. It is based on adaptive, self-organized compact genotypic encoding (SOCE) generating the phenotypic synaptic weights of the agent's neurocontroller. SOCE implements a parallel evolutionary search for neurocontroller solutions in a dynamically varying and reduced subspace of the original synaptic space. It leads to the emergence of compact successful neurocontrollers starting from large networks. The method can serve to estimate the network size needed to perform a given task, and to delineate the relative importance of the neurons composing the agent's controller network.

This article presents a new approach to the important challenge of localizing function in a neurocontroller. The approach is based on the basic functional contribution analysis (FCA) presented earlier, which assigns contribution values to the elements of the network, such that the ability to predict the network's performance in response to multi-unit lesions is maximized. These contribution values quantify the importance of each element to the tasks the agent performs. Here we present a generalization of the basic FCA to high-dimensional analysis, using high-order compound elements. Such elements are composed of conjunctions of simple elements. Their usage enables the explicit expression of sets of neurons or synapses whose contributions are interdependent, a prerequisite for localizing the function of complex neurocontrollers. High-dimensional FCA is shown to significantly improve on the accuracy of the basic analysis, to provide new insights concerning the main subsets of simple elements in the network that interact in a complex nonlinear manner, and to systematically reveal the types of interactions that characterize the evolved neurocontroller.