One of the problems with 'Good old fashioned AI' is the specificity of function and brittleness of its products. A glorified search algorithm capable of beating grand masters at chess is all very well, but it cannot be adapted easily to do much else. It will fall over if given unexpected input. This specificity is in part caused by an ignorance of the causal spread involved in generating complex behaviours. New approaches to the design of artificial systems show a greater awareness of this causal spread. This increased awareness is in part fuelled by the prevalence and effectiveness of natural systems that show large causal spread. Focusing mainly on genetic encoding and developmental processes, this essay is intended to be an explanation and exploration of the phenomenon of causal spread in natural and artificial systems. From this angle it will address the problem of how features can be elicited from natural systems that will be useful in the design of artificial systems. Natural and artificial examples will be used to show that causal spread must be embraced to produce non-trivial systems that are useful and robust.

Causal spread and exploitation

Cause and effect' is an oft-touted phrase used to imply a simple link between two events. If I hit a key on my keyboard, a letter appears on the screen. It is not really that simple though. In order for me to hit that key and for that letter to appear on the screen, a whole load of events occurred simultaneously. Influential factors could include the previous letter, biochemical activity in my muscles and brain, my frictional interaction with the keyboard, electrical impulses in the computer and so on. Casual spread offers a way of considering such complex causal networks rather than simple linear chains of events. A good place to start when discussing causal spread is somebody else's definition. Wheeler and Clark [1] define causal spread thus:

'We shall use the term causal spread to describe any situation in which an outcome is generated by a network of multiple, co contributing causal factors that extends across readily identifiable systemic boundaries.'

So there is a system within which some distinct event or set of events occurs. The causes for this are multiple. They are in a network, within which interactions can occur. The causes are potentially wide spread in source - inside and outside the system. The wide spread, crossing systemic boundaries nature of the causal network is important. Let us say there is an animal is in an environment, behaving. The dynamic and innate properties of the environment should be seen as equivalent in importance in inducing a behaviour (or in dictating the properties of that behaviour) to more traditional behavioural causes, such as intentionality on the part of the animal. Systems showing strong causal spread often have an exploitative feel. Thompson [2] used a genetic algorithm approach to evolve circuits on reconfigurable FPGA chips. When a functional circuit was evolved on the chip, it proved very difficult to figure out how it worked. Eventually it became clear that the GA had exploited irregularities in the chip itself such as to produce a design that used bizarre induction effects only available on that particular chip. Had the design (the 'behaviour') been evolved in simulation rather than on real hardware (the 'environment'), it would have been completely different. The relevance of causal spread in simulations is discussed later. Thompson's is a typical story from work using open-ended artificial evolution. Another example of exploitative causal spread is Karl Sims [3] work with simulated creatures, which he evolved in a simulated environment. Sims' creatures developed a strange, apparently impossible technique of movement that turned out to be an exploitation of a flaw in the simulated physics of his system.

A system that embraces causal spread

Webb's cricket robot [4] is an example of an artificial system inspired by a natural system with deliberate causal spread. From observations of the natural system, Webb was able to find the cross-systemic factors at play in a non-trivial behaviour and emulate them. The behaviour occurs as a result of the physical properties of several elements in the system - the nervous system, the environment and the body of the robot. This is in opposition to having a separate 'brain' processing input and sending instructions to its body. That is, the system is very much embedded in its environment; its behaviour is integrated with features therein. The boundaries between the implicated causal systems become blurred by the causal spread. This is in line with Brooks' [5] view that strong robot/ environment coupling is essential in designing robust robot systems.

Behaviour-stimulus feed back

An example of a view that does not appear to embrace causal spread is that described by Arkin [6] p48. Whilst looking into the usefulness of animal systems as models for robotic systems, Arkin discusses 'fixed action patterns', which are:

'...time extended response patterns, triggered by a stimulus but persisting for longer than the stimulus itself'

The stimulus is treated as a thing totally separated from the behaviour. From the causal spread perspective, the stimulus would be seen as integral to the behaviour. That is, the behaviour induces further stimulus and this stimulus feeds back into the behaviour. In this sense, the stimulus is not so much a static event, more a dynamic, interactive series of sense data that forms and is formed by the behaviour. Thus we are taken away from the view of animal brain as passive recipient of sense data towards an alternative view of an active, environmentally embedded system. There have been many studies of motion vision in the fly. [7] provides a mini review. Motion vision exemplifies a behaviour-stimulus feedback system in that without any behaviour, there is no stimulus.

In line with this view are Churchland et al [8] with their 'Critique of Pure Vision'. They discuss a non-orthodox future model for interactive visual perception in mammals where the visual input is not just passively received and sent up through hierarchies of classification. Rather the stimulation back projects into many areas of the brain including those associated with motor control. The theme is that the visually induced neural activity is sent back to the motor areas before it is fully classified. This is out of line with the assumptions of pure vision. Pure vision does not allow for these premature back projections of incompletely classified visual stimulus. In interactive vision, the back projection increases response speed and potential for perception-action feedback. Therefore, with what they call 'the integrated, whole-body character of visiomotor coordination', the animal plays an active part in generating its visual data.

Limiting the spread

An interesting applied angle on the concept of behavioural causal spread is the design of simulations for evaluation of evolving robot controllers. In order to evolve controllers for real robots it is necessary to evaluate the candidates of each generation. This can be done using real hardware, a method that is costly and time consuming, or it can be done in simulation. When designing a simulation it is necessary to reduce the computational expense and development time of the simulation such that it takes significantly less time to evaluate the candidates than it would do using real hardware. Jakobi [9] outlines a method for the construction of 'minimal simulations'. Minimal simulations are designed to be as fast running and easy to build as possible. Jakobi presents a seven-step procedure for minimal simulation construction. This procedure involves the terse specification of the desired behaviour, the fitness function and the relevant dynamics of the system. The most interesting step for this writing is number 2, 'Identify the real world base set'. This base set contains those real world features and processes that are relevant to the performance of the desired behaviour. The interactions between the robot and the real world base set as well as interactions within this base set must be modelled. In this paper an application of the minimal simulation approach to the evolution of an octopod robot controller is discussed. The behaviour is decomposed into 4 sub behaviours which produce non-trivial global behaviour and also make the fitness function easier to evaluate. The real world base set is focused on the interaction between the controller output and the body, i.e. how the body moves in response to controller signals. The external environment is pretty much ignored in terms of causal spread - it is reduced to a set of noisy sensor inputs. And for the robot body, the simulation simply makes a noisy estimate of the updated positions of the legs after controller output to their servos. Minimal simulation is a way of limiting causal spread such as to avoid exploitative results as seen in Thompson's [2] and Simms' [3] work. Jakobi designs these simulations with an awareness of the potential pitfalls of causal spread. The GA is not given a luxurious simulation to exploit. Instead it is given a randomly noisy impression of reality, a hard problem to solve. If it can find a solution in the face of this minimal, noisy world, that solution will transfer well across the 'reality gap' from simulation to real robot, as Jakobi calls it.

There are areas other than behavioural/ environmental couplings in which the lessons of causal spread can be applied. Wheeler and Clark [1] describe the mapping from genotype to phenotype in comparison to the mapping from neural states to behaviours. They are arguing against 'well individuated vehicles of internal representation' or genetic determinism. I will use this comparison to argue that the implementation of systems using developmental causal spread will yield benefits in a similar way to the implementation of systems exploiting behavioural (Webb)/ functional (Thompson) causal spread. I see the implementation of developmental causal spread as the introduction of developmental processes akin to those found in nature into artificial systems.

Developmental causal spread

Why let just the behaviour develop in tune with the environment? Why not let the body or controller of the system develop in the environment too? During the development of an organism, morphogenesis, the organism's cells interact in many ways that contribute to the resultant structures. To paraphrase Kauffman's [10] summary of the general framework of morphogenesis, cells may adhere to specific other cells. Cells may undergo spatially and temporally coordinated divisions that result in ordered spatial patterns. Cells may move and exert forces on each other. Cells may produce and respond to local and distant chemical signals, which affect their growth and morphology. The point to emphasise is that the genome of an organism does not explicitly describe the phenotype of that organism. In Wheeler and Clark's terms, the genome is not a complete representation of the organism's morphology. The various intercellular interactions mentioned above are integral factors in determining the morphology of the organism - the cause for this morphology is spread beyond the genome to the environment in which the development takes place. The use of the term 'environment' is extended to include the internal environment of the growing organism i.e. the other cells. Turing's [from 10] 1952 models suggest a process by which patterns spontaneously arise from uniformly distributed chemicals without any (genetic-type) control at all. Turing's models are extreme though - it does seem reasonable to allow the causal spread for neural/ body morphology to include genetic information!

Motivation for implementing developmental processes

The motivation for the promotion of such developmental processes is in part based on the existence proof that complex development occurs in nature. The structures thus formed are effective and the genotypes that contribute to their formation are relatively compact. The human genome could not possibly explicitly encode the morphology of the 1012 neurones in the brain, but it can encode a set of parameters and control circuits able to bound the developmental process. Development need not necessarily halt when the organism is 'complete'. If a human walks around with no shoes on, the skin on the bottom of its feet gets hard. If a pre-calibrated Mars rover robot arrives on the red planet to find that the gravity estimate is wrong and that it cannot walk, that is probably the end of its mission. On the other hand, the controller hardware could be left in an immature state. When it arrives in the working environment it matures coupled with this environment to a functional form. This is similar to behavioural learning but has a slightly different flavour. (The comparison of behavioural learning to morphological learning is beyond the scope of this essay)

Mind and body

There are two main areas in which the developmental process could be applied. First, the structure of the control system can undergo development. Second, the structure of the whole system could be developed. Research into the second variety [11] using real systems is limited by the problems of generating hardware that can change its physical structure. Some work has been done using modular, cube shaped robots that can connect to each other to make more complex forms [12]. The interactions between the robots were based on a self-organizing cellular automata-type system. They were able to form structures that allowed them to climb stairs higher than the individual robots. Other agent body morphogenesis research has been done by Delleart and Beer in simulation [13] Development occurs in a 2-D array of cells. Each cell has a copy of the genome, which is based on Kauffman's [10] random Boolean network. Such networks eventually settle into single states or limit cycles. The gene expression of each cell is dynamic and dependent on internal inter-gene interactions and external signals. When the gene expression pattern settles into a steady state, the cell divides. The daughter cells inherit the parental Boolean network state. The system was able to form an animat with morphology similar to an abstract chemotactic agent, i.e with bilateral symmetry and appropriately placed potential sensor structures. The most interesting part of this work was the introduction of an axonal growth system, using a cell adhesion molecule system related to that described by Boyan et al [14], mentioned later.

Returning to the first area for the application of developmental processes mentioned above, there have been many approaches to the genetic encoding of artificial neural networks. Not many of them seem to pay much attention to developmental causal spread. The neural architectures are often explicitly defined by the genome. This is clearly at odds with Kauffman's dogma of development mentioned above. A quick search through the developmental biology literature will find numerous complex neural development programs that have been, at least partially, elicited. For example Boyan et al [14] discuss axogenesis in the embryonic brain of the grasshopper. Pioneer neurones innovate pathways from the optic ganglia to the brain midline. These neurones strongly express cell adhesion molecule encoding genes, which allow them to form a scaffold on which the later neural structures can form. This system is clearly self-organising and not explicitly genetically encoded. In [15], Grapin-Botton et al explore the influence of non-neural cell sourced posteriorizing signals on Hox segmentation genes in the developing hindbrain of a quail/chick chimera system. Complex interacting chemical gradients are shown to control the differentiation of neural cells through the medium of genetic control circuitry.

Biological systems like these have provided inspiration for artificial neural development systems. Cangelosi, Parisi and Nolfi [16] implemented a system with axonal growth and neuronal cell division in a 3 layer 2-d plane. From a single initial egg, cell division and migration occurs for several cycles according to genetically encoded rules. Then axon growth cycles occur according to genetically encoded parameters such as branch angles and growth points (on the cell). If an axon contacts a neuron, a connection is made. During the cell division cycles a set of genetically encoded rules dictates the manner in which axonal growth parameters are passed from mother to daughter cell(s). This is akin to differentiation - successive generations of cells grow axons slightly differently. An evolutionary process is carried out to find egg cells capable of producing functional neural systems. The final position of the cells dictates the role they will specialise to have. If they are in the top out of the three planes, they will become motor neurones. If they are in the middle plane, they become hidden units. A bottom layer placed cell becomes a sensory or motivation neuron. The activity of the motivation neurons dictates the 'desires' of the animat. The researchers were able to evolve controllers capable of moving an animat around according to its motivational state. If it was hungry it moved to a randomly placed food area and stayed there long enough to reduce the activity of its hunger motivation. The genome of this system is still fairly complex though. The axonal growth is defined by explicit parameters in the genome. The genomic encoding scheme has been designed with its final operation very much in mind and as such has a top down feel to it. The task of the evolutionary algorithm is to find a solution that is kind of implicit in the system anyway. A more interesting approach would have been to allow simulated, axon growth influencing chemicals to diffuse from the cells such as to form a rich environment in which the axons could develop. Thus the causes for axon growth could be lifted from the genome and spread to the environment. The production of chemicals would be under genetic control, but this brings the system closer to the biological one in that the genetic control is made more implicit.

Conclusion

So it can be shown that behavioural causal spread is prevalent in natural systems; further that artificial systems embracing this phenomenon in terms of a tight behaviour-environment coupling are effective. Such systems tend to exploit the characteristics of their environment to their advantage rather than being limited by it. Examples include Webb [4] and Thompson [2]. In Thompson's case, a flaw in the FPGA chip environment is not a hindrance to the functioning of the system; rather it is integral to the functioning of the system. Can the concept be extended to justify the implementation of developmental processes in artificial systems, especially in artificial 'brains'? It cannot be denied that self-organising developmental processes are a vital part of many natural systems. They free the genome from the necessity of explicitly encoding the phenotype. This means the genome can be more compact since it defines a control system for a self-organisation process rather than every part of the organism. The complex neural structures found in animals could not possibly be explicitly encoded by a viably sized genome. As the functional demands placed on the control systems in artificial systems become greater, it seems inevitable that explicit genetic encoding will cease to suffice. Implementing developmental processes in artificial systems is not a trivial undertaking however. The search space explored by the GA in a system with self-organising properties is highly complex. Mutations in genes involved in different stages of development will have very different effects on the phenotype. Also reducing the influence of the genotype on the phenotype places more demands on an effective self-organisation system. Gruau et al [17] compared a direct genetic encoding for neural network controllers to Gruau's developmental cellular encoding system. They found that the developmental system took longer to find a solution but the solution was more compact and was found with less human intervention. The longer solution search is not surprising given that the search space was more complex. [17] shows that developmental systems are a viable option. The other research mentioned above, where artificial developmental systems have been successfully implemented further supports this. The future for artificial system behaviour is environmentally coupled neural architectures. The future for artificial system genotypes is compact with the causes for phenotypes spread beyond an explicit genetic encoding.

References

[1] Michael Wheeler and Andy Clark 'Genic Representation: Reconciling Content and Causal Complexity' - draft copy used

[2] Adrian Thompson 'An evolved circuit, intrinsic in silicon, entwined with physics' Proc. 1st Int. conf. On Evolvable Systems 1996

[3] Peter Coveney and Richard Highfield 'Frontiers of Complexity' pp342 - 343 1995

[4] Webb, B. "A Cricket Robot". Scientific American, December 1996, 62-67. (1996)

[5] Rodney Brooks 'Do elephants play chess?' in 'Designing autonomous agents' 1991

[6] Ronald C. Arkin 'Behaviour Based Robotics' p48 1998

[7] Egelhaarf and Borst 'Motion computation and visual orientation in flies' comp biochem physiol. Vol 104a no.4 pp 659-673) 1993

[8] Churchland, Ramachandran and Seinowski 'A critique of pure vision' in 'Large-scale neuronal theories of the brain' pp23-60 1994

[9] Nick Jakobi CSRP 497 Running Across the Reality Gap: Octopod Locomotion Evolved in a Minimal Simulation

[10] Stuart Kauffman 'The origins of order' 1993

[11] Philip K. Dick 'Second Variety' first pub. In UK 1989

[12] Hosowaka 'Autonomous modular robots' 1997

[13] F. Dellaert and R. Beer. 'Toward an evolvable model of development for autonomous agent synthesis' in R. A. Brooks and P. Maes, editors, Proceedings of the Fourth International Workshop on Artificial Life. The MIT Press/Bradford Books, Cambridge, MA, 1994.

[14] George Boyan, Stavros Therianos, J. Leslie D. Williams and Heinrich Reichert 'Axogenesis in the embryonic brain of the grasshopper Schistocerca gregaria' Development Volume 121 (1) pp75-86 1995

[15] Anne Grapin-Botton, Marie-Ange Bonnin, Michael Sieweke and Nicole M. 'Defined concentrations of a posteriorizing signal are critical for MafB/Kreisler segmental expression in the hindbrain' Development Volume 125 (7) pp 1173-1181 1998

[16] A. Cangelosi, D. Parisi, and S. Nolfi. 'Cell division and migration in a 'genotype' for neural networks'. Network: computation in neural systems, 1995

[17] Frederic Gruau, Darrell Whitley, Larry Pyeatt A Comparison between Cellular Encoding and Direct Encoding for Genetic Neural Networks