Evolving cognitive-behavioural dependencies in situated agents for behavioural robustness

Abstract

This article investigates the emergence of robust behaviour in agents with dynamically limited controllers (monostable agents), and compares their performance to less limited ones (bistable agents). ‘Dynamically limited’ here refers to a reduced quantity of steady states that an agent controller exhibits when it does not receive stimulus from the environment. Agents are evolved for categorical perception, a minimal cognitive task, and must correlate approaching or avoiding movements based on (two) different types of objects. Results indicate a significant tendency to better behavioural robustness by monostable in contrast to bistable agents in the presence of sensorimotor, mutational, and structural perturbations. Discussions here focus on a further dependence to coupled dynamics by the former agents to explain such a tendency.

Introduction

Studies in systems biology show that bacteria reach extreme robustness under harsh stress conditions (Alon et al., 1999). Robustness usually refers to the continuation of a function by a system (e.g. an organism) in the presence of perturbations (Kitano, 2007). It is assumed that bacteria can reach robustness by creating internal switches from one steady state to another to keep themselves functional, rather than trying to sustain a given state (Kitano, 2004). In such a dynamical interpretation, however, the roles of body (embodiment) (Ziemke, 2003) and spatio-temporal factors (situatedness) (Brooks, 1991) are considered despite the fact that bacteria are free moving organisms. This article proposes that behavioural robustness may well turn out to be a property of a particular organism-internal-control in its coupling with the environment, rather than a systemic property that is ‘ensured from inside’ (a common interpretation in current systems biology (Kitano, 2004)).

The growing consensus about the importance of internal-control, body, and environment couplings is still a minority view in several disciplines investigating biological robustness. These include cognitive psychology, neuroscience, a good part of Artificial Intelligence (AI) and robotics, and indeed several areas of biology. Minimal agents showing adaptive behaviour have, however, been extensively investigated through artificial evolution (e.g. Beer, 2003, Slocum et al., 2000, Williams et al., 2008), suggesting that adaptive and cognitive behaviour can be better understood from dynamically coupled interactions. Unfortunately, studies on robustness using bio-inspired systems are still in their infancy (Hubert et al., 2009).

This research aims to understand both (i) whether feedback from the actions that an agent produces in the environment is a decisive factor underlying behavioural robustness, and (ii) whether further dynamical complexity at neurocontroller level helps the emergence of robust behaviour. Without this feedback, it can be hypothesized that an agent will more easily be driven by perturbations to internal states that do not correlate with current environmental situations, and as such, it will produce non-appropriated categorical perception. For (ii), research in adaptive systems raises the question whether agents with further dynamical complexity at internal control can better cope with perturbations than dynamically simpler agents. In the former agents, internal dynamics could ideally transit between dynamical states, enabling agents to cope with the effects of perturbations (Kitano, 2007). In the latter agents, for instance, internal control could be rooted in transient dynamics around one attractor (internal state) (Buckley et al., 2008, Fine et al., 2007). Answers to question (ii) have a conceptual and practical interest for ethology and theoretical biology (Cliff, 1991), because they will provide a broad account of the simple dynamics and mechanisms underlying robustness (see (Hobbs et al., 1996, Teo, 2004)). In fact, robustness studies have not typically been addressed from the point of view of coupled dynamics (Silverman and Ikegami, 2010). Toward this aim, this article describes statistical, behavioural and dynamical analyses to answer questions (i) and (ii).

Work here contributes to this approach (coupled dynamics for robust behaviour) with experimental proofs and discussions from a computational perspective. The Evolutionary Robotics (ER) technique (Nolfi and Floreano, 2000) is used to give the right conditions during evolution for the emergence of dynamically limited control systems performing categorical perception (Slocum et al., 2000, Williams et al., 2008). Experiments in this article induce the emergence of agents that cannot exclusively rely on internal control for robust and adaptive behaviour, but they can exploit agent-environment coupling for the categorization task.

The next section introduces related works. The methods and experimental configurations are given in Section 3. Sections 4 Results, 5 Toward a Metastable Understanding of Multistability for Robustness, 6 Discussion the Effect of Agent's Perturbed Dynamics in Performance examine the results obtained and provide discussions to validate the hypothesis described.