Identification of Dynamical Structures in Artificial Brains: An Analysis of Boolean Network Controlled Robots

Abstract

Abstract. Automatic techniques for the design of artificial computational systems, such as control programs for robots, are currently achieving increasing attention within the AI community. A prominent case is the design of artificial neural network systems by means of search techniques, such as genetic algorithms. Frequently, the search calibrates not only the system parameters, but also its structure. This procedure has the advantage of reducing the bias introduced by the designer and makes it possible to explore new, innovative solutions. The drawback, though, is that the analysis of the resulting system might be extremely difficult and limited to few coarse-grained characteristics. In this paper, we consider the case of robots controlled by Boolean networks that are automatically designed by means of a training process based on local search. We propose to analyse these systems by a method that detects mesolevel dynamical structures. These structures are emerging patterns composed of elements that behave in a coherent way and loosely interact with the rest of the system. In general, this method can be used to detect functional clusters and emerging structures in nonlinear discrete dynamical systems. It is based on an extension of the notion of cluster index, which has been previously proposed by Edelman and Tononi to analyse biological neural systems. Our results show that our approach makes it possible to identify the computational core of a Boolean network which controls a robot.

Evolving Always-Critical Networks

“…The automatic design of BNs matching some target properties of biological cells, such as the distance between attractors or the capability of producing attractor landscapes with specific characteristics resembling differentiation trees, is the subject of more recent works [44][45][46]. Finally, it is worth mentioning that evolutionary algorithms and stochastic local search techniques in general have been also applied to design BNs which are capable of controlling robots [47,48], and the structural and dynamical properties of these BNs have been studied as well [49][50][51].…”

Complexity Measures: Open Questions and Novel Opportunities in the Automatic Design and Analysis of Robot Swarms

“…Apart from few preliminary investigations (Roli et al, 2013(Roli et al, , 2015(Roli et al, , 2018, there is still much room for application of information-theoretic and complexity measures in the analysis of robot behavior. Besides the use of such metrics during the design process (e.g., to assessing the complexity of the swarm or the individual robots), we envisage a real-world scenario in which the complexity level of a robot swarm is monitored in order to detect specific phases of its behavior (e.g., when a decision has to be collectively taken) and possible failures-e.g., when a large discrepancy between the expected and actual complexity of the swarm is observed it might be the case that some malfunctioning dynamics is taking place.…”

Attractor Landscape: A Bridge between Robotics and Synthetic Biology

“…Along this line are the experiments in Boolean network robotics [16][17][18][19]. A BN is a discrete-time, discrete-state dynamical system whose state is an N-tuple in 0, 1 N , (x 1 , … , x N ).…”

“…The algorithm employs as learning feedback a measure of the performance of the BN-controlled robot (in the following, BN-robot) on the task to perform, such as in evolutionary robotics [28]. For example, it was shown that a BN-robot can learn a composite mission, in which the first task is to perform phototaxis; then, after a sharp sound is perceived, the robot performs anti-phototaxis [16][17][18][19]. A dynamical systems' analysis shows that the behavior of the robot is mainly composed of three attractors: in the first the robot steadily rotates and in the second the robot goes straight.…”

“…Preliminary results in this direction have been achieved in controlling robots by means of BNs (GRN models introduced by Kauffman [13] and subsequently used to model important phenomena in biol-ogy [14,15]). The effectiveness of this approach was demonstrated through experiments on both simulated and real robots [16][17][18][19]. These experiments showed that BNs can be successfully used to control robots, and therefore that a nontrivial behavior can be attained by a system sharing some similarities with biological cells.…”