Recent artiﬁcial neural networks for machine learning have exploited transient dynamics around globally stable attractors, inspired by the properties of cortical microcolumns. Here we explore whether similarly constrained neural network controllers can be exploited for embodied, situated adaptive behaviour. We demonstrate that it is possible to evolve globally stable neurocontrollers containing a single basin of attraction, which nevertheless sustain multiple modes of behaviour. This is achieved by exploiting interaction between environmental input and transient dynamics. We present results that suggest that this globally stable regime may constitute an evolvable and dynamically rich subset of recurrent neural network conﬁgurations, especially in larger networks. We discuss the issue of scalability and the possibility that there may be alternative adaptive behaviour tasks that are more ‘attractor hungry’.
Buckley, C., Fine, P. Bullock, S. and Di Paolo, E. A. (2008). Monostable controllers for adaptive behaviour. In From Animats to Animals 10, The Tenth International Conference on the Simulation of Adaptive Behavior, Osaka, Japan, July 7-10, 2008.
This study presents an extended model of homeostatic adaptation designed to exploit the internal dynamics of a neural network in the absence of sensory input. In order to avoid typical convergence to asymptotic states under these conditions plastic changes in the network are induced in evolved neurocontrollers leading to a renewal of dynamics that may favour sensorimotor adaptation. Other measures are taken to avoid loss of internal variability (as caused, for instance, by synaptic strength saturation). The method allows the generation of reliable adaptation to morphological disruptions in a simple simulated vehicle using a homeostatic neurocontroller that has been selected to behave homeostatically while performing the desired behaviour but non-homeostatically in other circumstances. The performance is compared with simple homeostatic neural controllers that have only been selected for a positive link between internal and behavioural stability. The extended homeostatic networks perform much better and are more adaptive to morphological disruptions that have never been experienced before by the agents.
Iizuka, H. and Di Paolo, E. A. (2008). Extended homeostatic adaptation: Improving the link between internal and behavioural stability. In From Animats to Animals 10, The Tenth International Conference on the Simulation of Adaptive Behavior, Osaka, Japan, July 7-10, 2008
In psychology the ‘A not B’ error, whereby infants perseverate in reaching to the location where a toy was previously hidden after it has been moved to a new location, has been the subject of ﬁfty years research since it was ﬁrst identiﬁed by Piaget . This paper describes a novel implementation of the ‘A not B’ error paradigm which is used to test the notion that minimal systems evolutionary robotics modelling can be used to explore developmental process and to generate new hypotheses for test in natural experimental populations. The model demonstrates that agents controlled by plastic continuous time recurrent neural networks can perform the ‘A not B’ task and that homeostatic mediation of plasticity can produce perseverative error patterns similar to those observed in human infants. In addition, the model shows a developmental trend for the production of perseverative errors to reduce during development.
Wood, R. and Di Paolo, E. A. (2007). New models for old questions: Evolutionary robotics and the ‘A not B’ error. In Proceedings of the 9th European Conference on Artificial life ECAL 2007. Springer-Verlag.
Evolutionary robotics simulations can serve as a tool to clarify counterintuitive or dynamically complex aspects of sensorimotor behaviour. We present a series of simulations that has been conducted in order to aid the interpretation of ambiguous empirical data on human adaptation to delayed tactile feedback. Agents have been evolved to catch objects falling at diﬀerent velocities to investigate the behavioural impact that lengthening or shortening of sensory delays has on the strategies evolved. A detailed analysis of the evolved model agents leads to a number of hypotheses for the quantiﬁcation of the existing data, as well as to ideas for possible further empirical experiments. This study conﬁrms the utility of evolutionary robotics simulation in this kind of interdisciplinary endeavour.
Rohde, M. and Di Paolo, E. A. (2007). Adaptation to sensory delays. An evolutionary robotics model of an empirical study. Proceedings of the 9th European Conference on Artificial life ECAL 2007. Springer-Verlag.
This paper investigates the processes used by an evolved, embodied simulated agent to adapt to large disruptive changes in its sensor morphology, whilst maintaining performance in a phototaxis task. By avoiding the imposition of separate mechanisms for the fast sensorimotor dynamics and the relatively slow adaptive processes, we are able to comment on the forms of adaptivity which emerge within our Evolutionary Robotics framework. This brings about interesting notions regarding the relationship between diﬀerent timescales. We examine the dynamics of the network and ﬁnd diﬀerent reactive behaviours depending on the agent’s current sensor conﬁguration, but are only able to begin to explain the dynamics of the transitions between these states with reference to variables which exist in the agent’s environment, as well as within its neural network ‘brain’.
Fine, P., Di Paolo, E. A., Izquierdo, E. (2007). Adapting to your body. Proceedings of the 9th European Conference on Artificial life ECAL 2007. Springer-Verlag.