EPSRC Network on Evolvability in Biology & Software Systems

Evolvability, Genetics & Development in Natural and Constructed Systems: Abstracts of the EPSRC Evolvability Network Symposium

Tewin Bury Farm Hotel, Hertfordshire, England, UK
26-28 August 2003

University of Hertfordshire Computer Science Technical Report 389
C. L. Nehaniv, P. J. Bentley & S. Kumar (Editors)

Evolving Minimal

Evolving Minimal Developmental Systems for Piaget's Delayed Approach Task


Centre for Computational Neuroscience and Robotics
University of Sussex
01273 872952


The fundamental aim of robotics research is to implement adaptation; thus such work draws inspiration from knowledge of the processes which have served to adapt natural systems. Evolutionary robotics successfully exploits the principle of natural selection to find fit solutions to specific task environments. In the last decade significant progress has been made in adapting artificial agents and in using these experiments to further scientific understanding of natural evolution. Conversely, the exploration of ontogenetic adaptation has lagged behind and we are less able to adapt individuals than we are populations. One reason for this may be that early attempts to harness ontogenetic acquisition have focused on finding the ``right'' mechanisms to organise weights in a controller and neglected the evolutionary context which self-organisation occurs.

On this view, developmental processes are evolutionary artefacts but ontogeny is also a primary constraint on evolution. Ontogeny and evolution are fundamentally coupled in a relationship of mutual specification and we should expect significant interactions between adaptive processes despite their very different scales of operation. In this research we set out to explore some aspects of this relationship by evolving developmental systems.This approach uses evolutionary robotics as a base methodology for implementing simple, adaptive agents capable of on-line self-organisation in noisy task environments. By evolving embodied, situated systems in this way we can exploit the extant data on evolutionary models to guide our exploration of issues such as noise in ontogenetic adaptation and the trade-off between plasticity and robustness.

Some small steps in this direction have been made with two sets of experiments using Beer's minimal systems approach to evolve developmental processes. The first study was an extension of Di Paolo's acoustic approach task in which simple, simulated robots were required to locate and track a partner via sound signals. We extended Di Paolo's method to use plastic CTRNN type controllers seeking thus to obtain behavioural evidence of the effects of manipulation on individual developmental histories. The second series of experiments are based on Piaget's delayed manual search paradigm.Here, the task environment features two sound sources one of which emits a cue signal, after a short delay the robot is allowed to locate the source. During the approach phase both sources emit a low-level noisy signal, thus the robot is required to ``fix'' the location of the cue before making its approach.

This experiment has two broad aims, first to test how well this approach to modeling captures the dynamics of such a well-researched cognitive developmental phenomena. Our second aim focuses on what might be learned from a ``minimal'' A not B error, does the erroneous perseveration found with infant subjects have a ``cognitive'' component which this model cannot achieve? Does the error require disassociation between two systems sub-serving different aspects of sensorimotor coordination? Preliminary results indicate that the model has sufficiently complex dynamics to produce rich behavioural outcomes. Some evidence of perseverative A not B errors has been found and the model has demonstrated broad effects of the interaction between evolution and development.

It is important to note that this approach in no way seeks to explain the A not B error, no matter what error patterns are observed in these robots, we certainly cannot assume that infants and robots fail for the same reasons. However, a minimal developmental system which makes perseverative errors in the appropriate task environment can help us to discover what are the necessary and sufficient factors for the error to occur and thus guide further experimentation with infant subjects.