Better Living Through Chemistry: Evolving
GasNets for Robot Control
Phil Husbands
1,2
and Tom Smith
1
and Nick Jakobi
1,4
and Michael O’Shea
1,3
1
Centre for Computational Neuroscience and Robotics, University of Sussex, Brighton, UK
2
Scho ol of Cognitive and Computing Sciences, University of Sussex, Brighton, UK
3
Scho ol of Biological Sciences, University of Sussex, Brighton, UK
4
AnimatLab, Ecole Normale Superieure, Paris, France
{philh,toms}@cogs.susx.ac.uk, nick.jakobi@ens.fr, bafu7@central.sussex.ac.uk
Husbands, P., Smith, T.M.C., Jakobi, N. and O’Shea, M. Bett er Living Through Chemistry:
Evolving GasNets for Robot Control. Connection Science, 10(3-4):185-210. Carfax, 1998.
corresponding author:
Phil Husbands COGS University of Sussex Brighton BN1 9QH UK tel: +44 (0)1273 678556 Email:
philh@cogs.susx.ac.uk
keywords: ANN, diffusible modulator, evolutionary robotics, GasNet.
running heading: Evolving GasNets for Robots
1
Abstract
This paper introduces a new type of artificial neural network (GasNets) and shows that it is
possible to use evolutionary computing techniques to find robot controllers based on them. The
controllers are built from networks inspired by the modulatory effects of freely diffusing gases,
especially nitric oxide, in real neuronal networks. Evolutionary robotics techniques were used to
develop control networks and visual morphologies to enable a robot to achieve a target discrimina-
tion task under very noisy lighting conditions. A series of evolutionary runs with and without the
gas modulation active demonstrated that networks incorporating modulation by diffusing gases
evolved to produce successful controllers considerably faster than networks without this mecha-
nism. GasNets also consistently achieved evolutionary success in far fewer evaluations than were
needed when using more conventional connectionist style network s.
2
1 Introduction
1.1 Robots
Over the past decade there has been renewed interest within AI in building simple autonomous ‘crea-
tures’ as a way of investigating mechanisms underlying the generation of adaptive behaviour (Brooks,
1991a; Beer, 1990). The vast majority of researchers in this field use some form of artificial neural
network (ANN) as the basis of the ‘nervous system’ of their agents. These networks can be envisaged
as simple nodes connected together by directional wires along which signals flow. As has been pointed
out by various people (e.g. Brooks, 1994), advances in neuroscience have made it clear that the prop-
agation of action potentials, and the changing of synaptic connection strengths, is only a very small
part of the story of the brain (e.g Purves, 1997). This in turn means that connectionist style networks,
and even recurrent dynamical ones (Beer, 1990), are generally very different kinds of systems from
those that generate sophisticated adaptive behaviours in animals. Although our picture of biological
neuronal networks changes every few years, contemporary neuroscience can provide a rich source of
inspiration in devising alternative styles of artificial neural network (Brooks, 1991b). The core of
this paper is concerned with investigating abstractions of some of the extremely important chemical
mechanisms of nervous systems and incorporating them into control networks for simple autonomous
mobile robots.
1.2 Brains
Traditionally, chemical information flow in the brain has been thought to be mediated by messenger
molecules or neurotransmitters which are released by neurons at points of close apposition known as
synapses (Katz, 1969). Because most neurotransmitters are relative ly large and polar molecules (amino
acids, amines and peptides), they cannot diffuse through cell membranes and do not s pread far from
the release site. They are also rapidly inactivated by e nzymatic hydrolysis and by active re-uptake.
Together these features confine the spread of neurotransmitters very close to the points of release and
ensure that the transmitter action is transient. In other words, chemical synaptic transmission of
information operates essentially two-dimensionally (one in space and one in time). This conventional
interpretation is coupled to the idea that neurotransmitters cause either an increase or a decrease in
the electrical excitability of the target neuron. According to a traditional view of neurotransmission
therefore, chemical information transfer is limited to the points of connection between neurons and
neurotransmitters can simply be regarded as either excitatory or inhibitory.
3
In recent years two important discove ries have required a fundamental revision of this model. Firstly,
many neurotransmitters, pe rhaps the majority, cannot be simply c lassified as excitatory or inhibitory
(Hall, 1992). These messenger molecules are best regarded as ‘modulatory’ because among other things
they regulate or modulate the actions of conventional transmitters. Modulatory neurotransmitters are
also ‘indirect’ because they cause medium- and long-term changes in the properties of neurons by
altering the rate of synthesis of so called ‘second mes se nger’ molecules. By altering the properties of
proteins and even by changing the pattern of gene expression, these s econd messengers cause com-
plex cascades of events resulting in fundamental changes in the properties of neurons. In this way
modulatory transmitters greatly expand the diversity and the duration of actions mediated by the
chemicals released by neurons. Secondly, the discovery that the gas nitric oxide (NO) is a modula-
tory neurotransmitter has opened entirely unexpected dimensions in our thinking about how chemical
information is transmitted by neurons (Garthwaite et al., 1988; Gally, et al., 1990; H¨olscher, 1997).
Because NO is a very small and nonpolar molecule it diffuses isotropically within the brain regardless
of intervening cellular structures (Wood and Garthwaite, 1994). NO therefore violates some of the key
tenets of point-to-point chemical transmission and is the first known member of an entirely new class
of transmitter, the gaseous diffusable modulators.
NO is generated in the brain by specialised neurons that contain the neuronal isoform of the calcium
activated enzyme, nitric oxide synthase or nNOS (Bredt and Snyder, 1990). This enzyme catalyses the
synthesis of NO from the amino acid L-arginine and molecular oxygen. NO synthesis is triggered when
the calcium concentration in nNOS-containing neurons is elevated, either by electrical activity or by
the action of other modulatory neurotransmitters. The existence of a freely diffusing modulatory trans-
mitter suggests a radically different form of signalling in which the transmitter acts four-dimensionally
in space and time, affecting volumes of the brain containing many neurons and synapses (Bredt and
Snyder, 1992). The properties of NO that allow it to diffuse freely also prevent it from being stored
prior to release in membrane bound vesicles at the s ynapse, as are the conventional neurotransmitters.
This means that for NO to act as a signalling molecule, its release must be coupled directly to its
synthesis. Because the synthetic enzyme nNOS can be distributed throughout the neuron and NO
release does not require the synapse, NO can be generated and released by the whole neuron. NO
is therefore best regarded as a ‘non-synaptic’ transmitter whose actions moreover cannot be confined
to neighbouring neurons (Hartell, 1996; Park et al., 1998). NO cannot be classified conventionally as
excitatory or inhibitory, it is a modulatory transmitter which activates the synthesis of cyclic-GMP,
an important second messenger which regulates a wide variety of cellular processes in target neurons,
some of which underlie synaptic plasticity (H¨olscher, 1997).
The discovery of diffusible gaseous modulators in the brain clearly challenges simplistic connectionist
4
models of neural information processing (O’Shea et al., 1998). For example, it suggests a rich dive r-
sity of modulatory mechanisms with different temporal and spatial dynamics affect the properties of
neurons. Importantly, the discovery of diffusible modulators shows that neurons can interact and alter
one another’s properties even though they are not synaptically connected.
1.3 From Neuroscience to Control Systems
In this paper we have attempted to abstract some of these concepts and incorporate the elements of
gaseous transmission into a fundamentally new class of artificial neural network. Nodes in a spatially
distributed network can emit ‘gases’ which diffuse through the network. The ‘gases’ can modulate in-
trinsic properties of nodes and connections in a concentration dependent fashion. This paper describes
work where we have used this style of network to build control systems for autonomous mobile robots.
One of the new styles of AI to have emerged recently is Evolutionary Robotics (Cliff, Harvey and
Husbands, 1993; Nolfi et al., 1994; Floreano and Mondada, 1994; Yamauchi and Beer, 1994; Husbands
and Meyer, 1998). The evolutionary process, based on a genetic algorithm (Holland, 1975), involves
evaluating, over many generations, whole populations of control systems specified by artificial geno-
type s. These are interbred using a Darwinian scheme in which the fittest individuals are most likely to
produce offspring. Fitness is measured in terms of how good a robot’s behaviour is according to some
evaluation criterion. This selectionist approach is particularly suited to the exploration of classes of
control networks involving many parameters and whose properties are difficult to predict in advance.
The type of networks introduced in this paper are of that nature and have been investigated using
evolutionary robotic techniques.
A word of warning: the focus of this paper is on ANNs using computationally efficient loose abstrac-
tions of biological phenomena; there is no modelling involved. However, for brevity and convenience,
biological terminology is used frequently – it should be taken as analogy only. Having said that, the
kind of work described in this paper can potentially have a useful relationship with more explicit mod-
elling studies (Philippides et al., 1998; Gally et al., 1990). Of course it should also be stressed that
the physical language used in this paper (gas, diffusion etc .) is again only analogy. The networks are
actually abstract discrete dynamical systems implemented as a C program. However, we feel that a
profitable way to think of these systems are as computational simulations of physical devices. This
particularly because they are used to control physical devices engaged in activities where space and
time are highly pertinent. For instance, the coupling between the robot’s visual sensors and the control
networks means that there is a direct relationship between the (modelled) spatial prop e rties of the net-
works – crucial to the operation of the ‘diffusion’ processes – and the (actual) spatial properties of the
5
