Analysis of Evolved Sensory-motor Controllers Analysis of Evolved Sensory-motor Controllers

We present results from the concurrent evolution of visual sensing morpholo gies and sensory motor controller networks for visually guided robots In this paper we analyse two of many networks which result from using incremental evolution with variable length genotypes The two networks come from separate populations evolved using a common tness function The observable behaviours of the two robots are very similar and close to the optimal behaviour However the underlying sensing morphologies and sensory motor controllers are strikingly di erent This is a case of convergent evolution at the behavioural level coupled with divergent evolution at the morphological level The action of the evolved networks is described We discuss the process of analysing evolved arti cial networks a process which bears many similarities to analysing biological nervous systems in the eld of neuroethology Introduction As part of our ongoing work in using genetic algorithms to develop neural networks which act as controllers for visually guided robots we have analysed the nal evolved networks in order to identify how they work This is an essential step in moving away from the treatment of arti cially evolved neural networks as magical black boxes The mathematics of our particular style of network are such that it would be di cult or impossible to derive closed form equations describing the action of the networks Instead we analyse our networks using techniques analogous to those used in the study of biological sensory motor neural systems In trying to understand how our arti cially evolved networks generate behaviours in the robot we are performing a task directly analogous to the task faced by biological scientists in the eld of neuroethology Neu roethology is the study of the neural mechanisms underlying the generation of a creature s behaviour see e g For further details of the link between neuroethology and arti cial neural network research see We view the networks we evolve as continuous dynamical systems rather than as computational devices transforming between representations inputs to the system might perturb the trajectory of the network in state space so it enters a di erent state which For example the transfer functions used in our model neurons are all nonlinear with discontinuities in the rst derivative and non Gaussian noise is introduced at a number of points in the sensory motor system might be interpreted by an external observer as a new behaviour We nd this perspective less encumbering than the traditional computational perspective and also less amenable to the use of potentially misleading intentional language see e g for further discussion of the bene ts of adopting a dynamical systems perspective Most of this paper deals with analysing two networks from separate populations each evolved to perform the same task We demonstrate that although the nal observed behaviour from the two networks is very similar the underlying mechanisms are remark ably distinct the two populations converged at the behavioural level while maintaining distinct sensory motor morphologies The primary focus of this paper is on analysing networks resulting from the evolution ary processes The text refers the readers to past papers for further details of the genetic encoding the genetic algorithm employed and description of the vision system Never theless Section o ers a brief overview of most of the important details Following that Section describes our experimental regime and provides analysis the two networks Finally Section discusses the implications of our work Background Rationale The rationale for our work and some early results have been discussed elsewhere The notes below present a brief summary of the important concepts In common with a growing number of other researchers we believe that the generation of adaptive behaviour should form the primary focus for research into cognitive systems By adaptive behaviour we mean behaviour which is selected to increase the chances that a situated agent can survive in an environment which is noisy dynamic hostile and uncertain Almost all animals in the natural world exhibit some form of adaptive behaviour and there is increasing interest in the creation of arti cial systems which are capable of acting in an adaptive manner The arti cial systems are commonly either simulated virtual agents or real robots Our work to date has involved using arti cial evolution on populations of simulated robots The simulations involve a model of a real robot built at Sussex and the simulated vision employs advanced computer graphics techniques Work is currently underway on the construction of specialised robotic equipment which eliminates the need for simulating perception and action while still allowing the use of arti cial evolution see for further details For reasons given in we are approaching the task of creating arti cial agents that exhibit adaptive behaviour in accordance with the following set of beliefs Neural network processors are likely to be most useful in building controllers for agents that exhibit adaptive behaviour Namely ray tracing with antialiasing via sixteen fold supersampling see e g for details of such techniques Manual design of such networks is likely to become prohibitively di cult as in creasingly complex or sophisticated behaviours are required Rather than design by hand we are employing arti cial evolution techniques based on Harvey s saga variable length genotype methods Almost all adaptive behaviours bene t from distal i e long range sensory infor mation While there is an established body of successful work studying robots with only tactile sensing i e mechanical whiskers the proximal nature and low dimensionality of the robot s sensors constrain it to relatively primitive bumping and feeling behaviours such as wall following For demonstration of our methods working with only proximal sensors see A primary means of gathering distal sensory information is by use of visual sensing so we believe visually guided agents should be studied from as early a stage as possible While we could impose on our robot some visual sensors with xed properties we advocate in common with Brooks the concurrent evolution of visual sensor morphology and the control networks separating morphology from control is a measure which is di cult to justify from an evolutionary perspective and poten tially misleading For reasons of parsimony studies of visually guided agents should commence by examining minimal systems The work reported on here involves robots using very simple low resolution devices coupled to small networks It is our intention to work towards more complex i e higher resolution systems Furthermore because we intend to transfer our results from simulated robots to the real robot on which the simulation is based we constrain evolution such that the evolved designs could realistically be built from discrete components and operate in real time In e ect our intention is to evolve a speci cation for a robot with electronic compound eyes c f Details In accordance with the last item in the above list our current studies have addressed evolving visually guided robots with just two photoreceptors i e two pixels in the input images The direction of view of the photoreceptors and their acceptance angles are under evolutionary control it in this sense that the visual morphology is concurrently evolved along with the controller network For full details of the genetic encoding for both the control networks and the visual system see Because there are only two photoreceptors we can only expect to evolve robots which exhibit relatively simple behaviours Nevertheless we have concentrated on evolving robots which perform tasks that would be di cult or impossible using only tactile information Physically the Sussex robot is cylindrical it has a circular bottom plate on which the motors and wheels are mounted and a circular top plate where a notebook computer is situated the computer simulates the control networks The robot has three wheels arranged to give tripod stability At the front are two independent drive wheels each capable of rotating at one of ve speeds full on half on o half reverse full reverse The rear wheel is a large ball bearing freewheel castor The robot is equipped with tactile sensors giving a six bit input vector it has four radially oriented binary whiskers and binary bumper bars at front and rear For illustration see The simulated robots are accurate models of such a vehicle with the addition of visual sensors While our early tactile only work involved the robot roving around cluttered o ce like environments all the visually guided tasks have been set in a closed circular arena The arena has black walls while the oor and ceiling are white There are no obstacles the arena contains only the robot The visual input from each of the robot s photoreceptors at any particular moment in time depends on the robot s visual morphology and the position and orientation of the robot in the arena Essentially the population of robots has to evolve to correlate the visual input with its position in the world so as to satisfy whatever tness evaluation we impose on the robot s behaviours As was demonstrated in visual guidance emerges without explicit reference to vision in the evaluation process In the early stages of evolution the tactile sensors can be useful in helping correlate visual input with the robot s position However as will be demonstrated below later generations typically tend to rely only on visual information Networks and the Neuron Model The controller networks are continuous dynamical systems built from model neurons i e processing units which can have asymmetric and recurrent connectivities Acti vation values all real numbers in the range are transmitted between units along the connections all of which have a weight of one and impose a unit time delay in transmission Fully asynchronous processing is simulated by ne time slice approxima tion techniques with random variation in time cycling on each unit to counter periodic e ects The neuron model has separate channels for excitation and inhibition A schematic of the operations for one unit is shown in Figure The inhibition channels operate as a veto or grounding mechanism if a unit receives any inhibitory input its exci tatory output is reduced to zero but it can still inhibit other units Excitatory input from sensors or other units is summed if this sum exceeds a speci ed inhibitory output threshold the unit produces an inhibitory output Independently the sum of excitatory inputs has uniform noise distribution n where n is a real number added and is then passed through an excitation transfer function the result of which forms the excitatory output for that unit so long as the unit has not been inhibited For further details of the excitation transfer function see We have found that this neuron model is su ciently sophisticated that there has been no need to introduce variable connection weights or variable delays for controllers based on the minimal visual systems studied so far Nevertheless we are actively investigating the use of placing such parameters within evolutionary control