Environment-driven distributed evolutionary adaptation in a population of autonomous robotic agents

Abstract

This paper is concerned with a fixed-size population of autonomous agents facing unknown, possibly changing, environments. The motivation is to design an embodied evolutionary algorithm that can cope with the implicit fitness function hidden in the environment so as to provide adaptation in the long run at the level of the population. The proposed algorithm, termed mEDEA, is shown to be both efficient in unknown environments and robust to abrupt and unpredicted changes in the environment. The emergence of consensus towards specific behavioural strategies is examined, with a particular focus on algorithmic stability. Finally, a real world implementation of the algorithm is decribed with a population of 20 real-world e-puck robots.

Embodied artificial life at an impasse can evolutionary robotics methods be scaled?

“…In the long run, this is still likely to be the case, but has been less of an issue than might have been expected. Some of the more complex tasks discussed in the literature have successfully employed aggregate or nearly aggregate fitness functions [2,56,59,61]. The ramifications of this are of interest and provide an additional point of support for the conjecture presented in the previous subsection of this paper: the set of benchmark tasks used in ER contain solutions that are less complex than previously supposed.…”

“…Recently there have been additional advances in ER in terms of complexity of behavior [2,30,32,33,[59][60][61][62]. We will look at some of these in more detail in later sections of this paper.…”

“…Currently more complex foraging tasks are a focus in ER. Examples include the sequential ball collection and deposition task summarized in the introduction [2] and a large-scale 100-robot foraging task evolved with an environmentally situated GA [61].…”

“…Some evolved autonomous agent behaviors seem to display a degree of scalable complexity [2,56,60,61], but this may partially reflect a functional overlap between synthesis and optimization (adaptation) processes at the lower end of task complexity. Here we suggest that controller configurations within a certain radius of a solution or optimum (in terms of A. L. Nelson in proc.…”

From evolutionary computation to the evolution of things

“…For example, Filipic et al evolved control strategies for operating container cranes using a physical crane to determine fitness values 37 . The evolution of controllers is also possible in situ -for example, in a population of robots during, and not just before, their operational period 38,39 . Evolutionary robotics 40 is an especially challenging application area because of two additional issues that other branches of EC do not face: the very weak and noisy link between controllable design details and the target feature(s); and the great variety of conditions under which a solution should perform well.…”