Humans are rather funny things, we often tend to imagine that we are so ‘special’, so divorced by our supposed ‘intelligence’ from the influences of the ‘natural world’ and so unique in our ‘abstracting’ abilities. We have this persistent delusion, evident since ancient Greek times, that we are ‘rational’, that we can behave as ‘disinterested observers’ of our world, which manifests in AI thought today in a belief that, in a like manner, we can ‘design’, God like, from afar, our replacements, those ‘super-robots’ that will do everything that we can imagine doing, but in much ‘better’ ways than we can achieve, and yet can avoid doing anything ‘nasty’, i.e. can overcome our many human failings - obeying, I suppose, in the process, Asimov’s three ‘laws of robotics’. Such human naïveté proves, in fact, to be quite amusing, at least to those of us ‘schooled’ in AI history. When we look at the aspirations and the expectations of our early ‘pioneers’, and compare them to the actual reality of today, then we must, it seems, re-discover the meaning of the word ‘humility’. Enthusiasm, good as it may be, needs to be moderated with a touch of ‘common sense’, and if our current ways of doing things in our AI world don’t really work as we had hoped, then perhaps it is time to try something different (Lucas, C., 1999)?
The traditional AI approach, being a top-down process, echoes the general behaviours seen in our world today, which attempt to centralise power and to have one ‘designer’ (or a small group of them) create a ‘robot’ or ‘system’ based upon a specification of some sort or another. In other words we need to first decide what ‘someone’ wants to achieve and then to implement or to impose a way of arriving there. Unfortunately the success of this method has been rather slight in practice, we still don’t know enough about the basis of intelligence to ‘design it’ effectively - especially if we wish to mimic what humans actually do ‘well’, rather than what they do ’badly’ (which we can, rather irrelevantly, just ‘manage’, it seems, to do artificially!).
As a way of overcoming the limitations of this ‘outsider’ method, an alternative has been proposed, i.e. the subsumption architecture of (Brooks, R., 1990). Here we concentrate our attention on a number of relatively simple operations, for example ‘moving forward’ or ‘turning’. Each of these is implemented in an autonomous module and these are then arranged into a layered hierarchy, with the most ‘primitive’ at the bottom. Each module can then inhibit the higher (more valued) modules, whenever their lower function is needed or is ‘necessary’, in other words we have ‘priority interrupts’. In this way we can avoid the need to plan out exactly what should happen in every possible scenario, instead we leave it to the environmental feedback to ‘select’ (evolutionary fashion) which ‘specific’ module needs to be operational at any time and for how long. Whilst this has proved quite successful, allowing the emergence of unexpected behaviours that rather look ‘intentional’ - although these systems are teleonomic not teleological, (so it is only the ‘observer’ that imputes ‘intention’ to them), we find that significant limitations still exist, e.g. in the need for explicitly ‘designed’ operations.
Given that this architecture relies on the environment, but still requires that each module be ‘hand-crafted’, can we go yet a stage further and dispense with that ‘designer’ stage, allowing feedback itself to ‘sculpt’ the robot entirely, i.e. in the same manner as is thought to happen in evolutionary biology? In traditional neo-Darwinian evolution (Futuyma, D.J., 1986) we rely on genetic mutations to generate the variety on which selection then acts. This however is a very long term process, it took billions of years before single celled organisms (e.g. bacteria, which reproduce every 20 minutes or so) achieved multicellularity, and a billion more years were needed before ‘intelligent’ creatures arrived as a reality. It is most unlikely that any of us in the AI community could live long enough to achieve even a basic prototype!
Fortunately we can take advantage of some short cuts. One of these acts by using computers, in order to model the multi-generational ‘phylogenetic’ evolution scenario just outlined. We can, using high speed computers and a technique called ‘genetic algorithms’ (Holland, J., 1992; Lucas, C., 2000a), operate at speeds of hundreds of generations per second - increasing as computers become faster. In this way we are able to ‘evolve’ some structure in a reasonable length of time, but what this happens to contain depends very much upon the constraints we apply to the system. These constraints mimic the ‘selection’ phase of natural evolution, but what should they be? Once again, in order to get what we want, so that we can arrange to ‘select’ for it, we need to know what it is in advance - so how can we avoid designing our ‘fitness function’ appropriately, but so deterministically, and then suffering the problem that evolution stops altogether once the population has ‘converged’ on our fitness optimum?
One way around this new problem is to use coevolution. This means that we let one ‘organism’ act as the fitness function for the other, and vice-versa. By using this sort of technique we can indeed generate improved functionality, for example in deriving efficient ‘sort programs’ where the list of numbers to be ‘sorted’ also evolves in difficulty (Hillis, W.D., 1991). But we have to start somewhere! We must create at least ‘prototypes’ of program and list, before anything at all can happen. So we seem yet again to be forced back to stage one - deliberate design. We can minimise the extent of this to some degree by simply designing-in the ability to adapt to unpredictable environments, as in ‘reinforcement learning’ (Ackley, D. H. & Littman, M. L., 1992), but the basic ‘intelligence’ to so act must still be ‘crafted’ by our ‘outsider’.
But if we can’t find a way to get over this problem then how can ‘nature’ possibly do so, do we need a ‘God’ as some seem to think? Here we come to the crux of the matter, in considering a process that has been left out of standard neo-Darwinian evolution, that of ‘development’ or ‘ontogeny’. Every human ‘robot’ starts life as a single cell (like our bacteria), this then grows into an embryo and thus eventually is born as a (more or less) functioning human child. But that is not the end of the matter, life experiences then develop, initially, the brain (our neural network wiring) and later (in interactions with others) our mind or ‘personality’, i.e. our behavioural range. It is this latter state we hope to duplicate in AI, so looking at how this process is understood in developmental biology should prove to be a useful indicator as to how we might achieve, artificially, much the same result, and without the need for any form of external ‘intelligent designer’.
Unfortunately it isn’t understood very well at all as yet, at least in overall terms! Embryogenesis, as it is called (Bard, J., 1990; Slack, J.M., 1991; Wolpert, L., 1998), consists of three stages. Firstly ‘growth’ (the duplication of cells by asexual or ‘mitotic’ reproduction), secondly ‘differentiation’ (the spliting of cells into different tissue types) and thirdly ‘morphogenesis’ (the creation of form or structure). The first stage expands possibility space, for every doubling of the number of cells we have a combinatorial explosion, e.g. for just 64 cells we have over 1089 permutations (64!), and this escalates rapidly during the growth process. But let us look more closely now at the second stage, after all, these permutations only become really interesting if the cells can be distinguished from each other. What is known about this is that there are many genetic regulatory networks involved, each ‘cell type’ (and there are hundreds) activates a different set of genes. These regulatory ‘controls’ do not however operate in a linear hierarchy, 1:N, in the way that we often regard human control networks, nor do they have fixed functions, independently maintained, as we often consider in AI ‘subroutines’.
What happens is that they are arranged into a web, in N:M fashion, where each ‘gene’ inter-links with many others, switching them on and off and being controlled in a like manner (Lucas, C., 2004b). This is a ‘circular feedback’ form of causality, called ‘polygeny’ and ‘pleiotropy’, which actually operates on several different levels. The same processes of activations and inhibitions also take place between cells (the third stage of embryogenesis), and this includes the inter-neuron connections in our mind. We can even, if we wish, go beyond the mind as an entity in itself and venture out into the wider world, and we shall see exactly the same phenomenon, in society and in ecology both, i.e. a ‘systems’ viewpoint of some type is applicable in all areas of our world.
What is common then to all these processes and levels? Well, it is the idea of ‘self-organization’ which is our focus in this article, which we can define like this (Lucas, C., 2004d):
“The essence of self-organization is that system structure often appears without explicit pressure or involvement from outside the system. In other words, the constraints on form (i.e. organization) of interest to us are internal to the system, resulting from the interactions among the components and usually independent of the physical nature of those components. The organization can evolve in either time or space, maintain a stable form or show transient phenomena. General resource flows within self-organized systems are expected (dissipation), although not critical to the concept itself.
The field of self-organization seeks general rules about the growth and evolution of systemic structure, the forms it might take, and finally methods that predict the future organization that will result from changes made to the underlying components. The results are expected to be applicable to all other systems exhibiting similar network characteristics.”
In other words, we now have a method of generating demonstrable structure for ‘free’ (Heylighen, F., 1999), of getting over (maybe) our primary ‘design’ problem. To see just how significant this is, let us add a few numbers. Suppose we have 10,000 randomly connected, 2 input, logic gates (what is called a ‘random Boolean network’ of size N - Lucas, C., 2002b), in other words the number of different possibilities is 210,000 - a staggering number. On average (with considerable variance) we would expect any area of the self-organized system to only visit (cycle through) 100 (square root of N) different dynamical states - this is less than 27, and there will be (again, with considerable variance) only about 100 disjoint areas of activity, 27 again (root N). So the ratio of initial ‘disorder’ to final ‘order’ is a massive 29986 or so ! This is not quite such a ‘magic’ solution as we may wish however, we still must have ‘something’ initially to work with (the gates here), but this proves to be our second ‘short-cut’. If the ‘lower levels’ also emerged (as we think) in this way, then we can ‘cut to the chase’ as it were, and ignore how they actually got there. We thus can start off with ‘parts’ of appropriate types for our AI purposes, called ‘agents’, and let a collection of these coevolve and self-organize, bottom-up, to meet our needs. All we have to do then is to sit back and watch it happen...
Studies of the dynamics of such scenarios (Kauffman, S., 1995) show that three general results are possible. In the first, the agents are insufficiently connected (too cool), they don’t interact much at all, so the system quickly settles into a fixed state, we have convergence to a ‘static’ result (akin to the traditional single analytical ‘solution’ in science). In the second the agents are highly connected (too hot) and each affects many others constantly, here the system cannot settle, it is always ‘perturbed’ (Lucas, C., 2000b) and exhibits a ‘chaotic’ behaviour (those ‘insoluble’ systems usually ignored in science). In the third state, which I call ‘Type 4 Complexity’ (Lucas, C., 1999), we have a (just right) behaviour that modularises the system, with some sets of agents proving to be static, some chaotic, and some dynamic - many of which will swap places over time as the system evolves between the possible (semi-stable or multistable) ‘attractors’ (Lucas, C., 2002a). In this scenario we find that the maximum ‘fitness’ can be achieved, the best overall performance is within reach. For larger systems, the dynamics will achieve a ‘fractal’ or ‘power law’ spread of properties, called ‘Self-Organized Criticality’ or ‘Edge-of-Chaos’, (Bak, P., 1996; Lucas, C., 2000b), which can give a somewhat emergent multi-layered or hierarchical structure, with inherent cooperative behaviour between the parts becoming apparent (Ünsal, C., 1993).
This idea of appropriate ‘connectivity’ is proving to be highly important in many areas of our world, from the social ones related to ‘anarchy’, ‘democracy’ and ‘totalitarianism’, via medical ones related to ‘epidemics’, through to ecological ones related to ‘diversity’ and freedom of combination, not to mention the physics or mathematical ones related to ‘spin glasses’ or ‘percolation’. By arranging connectivity suitably we can enable our important self-organizational processes. This is the ‘communications’ aspect of agent interaction, but two other aspects also need to be included here if we are to achieve success in our self-organizational scenarios. The first is appropriate size, relating to decentralisation. Systems must be small enough to be self-contained - if they are too big then the inertia of ‘bureaucracy’ inhibits the recognition of any ‘improvement’; but they should not be too small either - else they will have insufficient ‘variety’ with which to make improvements. The other aspect is ‘stress’, a desire or a need for improvement. Again if this proves to be too high the system will disintegrate, we will have rapid breakdown; but if the stress is too low then the ‘status-quo’ cannot be overcome and a static state will persist. Given that these ‘middle-way’ conditions are met, then self-organization should occur and the system will generate our required ‘novelty’ or emergence.
Although this explanation sounds very glib and easy, in practice there are a number of problems, for example in the (very visible) social and environmental destructiveness that we can see around us, resulting from the unfettered individualism driving the self-organization (‘invisible hand’) of the (over-stressed) ‘free market’. Experiments using Multi-Agent Systems (MAS), which operate using these ideas of self-organization, have also not so far achieved those higher levels of structure that we so desire and expect (seen in nature in the progression atoms-molecules-cells-organisms-societies-ecologies), and which are commonly to be found in the behaviour of ‘swarms’, for example, in insect societies (Bonabeau, E. et al., 2002), where ‘stigmergy’ (environmentally mediated communication) also has an important effect (Holland, O. & Melhuish, C., 1999). These current failures may well be because of the assumptions embedded within the agent structures typically used. In so many current systems, there is an inherent ‘competitiveness’ - echoing the belief behind the phrase ‘survival of the fittest’ often employed by Darwinists (and capitalists). Yet let us consider cellular development once more, what would be the effects of such ‘competition’? The answer seems pretty clear, it is the same as what happens when we suffer from ‘cancer‘ - the competition from ‘rogue’ cells eventually destroys the host. Thus it is not ‘competition’ that we need, but ‘cooperation’. We need to find a way for the agents to ‘work together’, since only in this way can organisms (and/or societies) function and persist at ‘higher levels’.
The principle we are looking for, which we wish to employ within our self-organizing systems, is called ‘synergy’ (Corning, P., 1995; Lucas, C., 2004c). Here, when two or more agents come together, a new ‘functionality’ arises, they gain combined powers greater than the sum of their separate powers, often illustrated with the phrase “the whole is greater than the sum of the parts”. But how can this possibly work? In essence, by a form of combinatorial trial and error - in which, in the processes of interaction, this new higher level functionality arises. Thus there is initially a diversity and an ongoing novelty, as seen in the pairwise encounters of the heterogeneous agents, but in some way these agents then ‘associate’. This, like the sexual crossover experienced in evolution, allows new ‘building blocks’ to arise, new combinations of functions which may, perhaps, operate in an entirely new way - we have potential ‘emergence’ (Lucas, C., 2004b). In the operation of the typical MAS the agents interact and learn (at least partly) in a random fashion, and their individual behaviours change, but they do so not only as a result of their own experiences - we find instead that the ‘higher level’ places new constraints upon them (Epstein, J.M. and Axtell, R., 1996). These constraints, called ‘downward causation’ (Campbell, D.T., 1974; Lucas, C., 2004b) add new values to the system, new environmental relevances at a ‘group’ level that imply selective forces beyond those of the individual. Although such ideas have been resisted within biology for some decades, only recently making a comeback (Wilson, D.S. & Sober, E., 1994), they do prove to be valid from both computing and complexity science perspectives, e.g. (Sloman, A. & Chrisley, R., 2003).
For this to prove useful however there must be a possibility of a dynamic from the agents to the ‘end result’, a way of searching current ‘state space’ (Lucas, C., 2002a), and expanding it in ways that enables such new functionality. Yet this aspect of emergence is very much under researched so far, we have very little idea as yet as to how we can arrange systems such that anything predictable will emerge, let alone to achieve what we, ideally, would wish to see. This is perhaps the greatest challenge to be met in the future by the complexity science community. But given this limitation, can the ideas we have outlined contribute already in any way to current robotic research ? We shall see that they can and they do.
An implied embryogenesis perspective provides some, highly scaleable, advantages for robot designers (Bentley, P.J. & Kumar, S., 1999). These include ‘adaptability’, the ability to respond to context (Quick, T. et al., 1999); ‘compactness’, the ability to code large structures in an efficient form; and ‘repetitiveness’, the ability to reuse the same structures or subroutines for many different functions. By using these techniques, in for example the evolutionary design of neural networks (Astor, J.C & Adami, C., 2000), we can evolve functional robot controllers (Jacobi, N., 1995), which potentially can interface with humans (Kanada, Y. & Hirokawa, M., 1994). Self-organization is also a useful biological technique which can be used for evolving robotic functionality (Nolfi, F. & Floreano, D., 2000; Kim, D.H., 2004), thus by combining the two perspectives (low-level agent development and higher level agent interactions) we may, advantageously, enable an ‘embodied’ form of ‘autopoietic’ (Lucas, C., 2004a) emergence, a coevolution of situation and actor. Note that we have two opposing drives here, the first (embryogenesis) expands state space, it add new possibilities, new options or combinations to the mix, the second (self-organization) reduces this diversity, it selects from those many options only those possibilities than can persist, the functionally stable states of the system. It moves from the system’s starting points in state space to its ‘attractors’. But these should not be viewed in isolation, they are only stable in terms of the current environment, if the context changes then that stability can be lost and another, alternative, stable state must then be found. This is, in fact, what we mean by learning or ‘epigenesis’, the move from one stable state in a certain context to another stable state in a different context. If our robot cannot do this, if it fails to ‘adapt’, becoming what we term ‘fragile’, then it has insufficient options (‘requisite variety’ in cybernetic terminology) to cope with the diversity of its environmental perturbations.
Designing robots that can overcome this tendency to be highly domain restricted has been a major headache in AI history, so how can our new perspective help? If we are to ‘grow’ some form of robot from scratch then four aspects are necessary. Firstly we must have a part (or a set of different parts) which can increase in number, secondly those parts must be able to associate (communicate and/or stick together) in some way such that they can form aggregates (some equivalent to ‘cell-adhesion’ or ‘morphoregulatory’ molecules - Edelman, G., 1992), thirdly they must have the freedom to self-organize, (i.e. to change their configurations and communications dynamically), and finally we need to allow for environmental influences to be able to trigger these reorganizations (allowing adaptability). To see the latter two aspects in a different light, self-organization restricts (‘canalizes’) the possibilities open to the system, it is a form of internal selection. The environment puts stress or bias on the system to achieve a viable function, causing it to escape poor attractors and flip to better ones, it is a form of external selection (if our system can’t so adapt it simply dies). But as we add new units to the mix, as we ‘grow’ the self-organizing system (Fritzke, B,. 1996), perhaps creating a 3D ‘morphology’ e.g. (Eggenberger, P., 1997), then we both add to and reconfigure its attractors, so that in this way we can increase the ‘requisite variety’ until our system can, in fact, cope with the target environment. This is similar to the way in which we make additional synaptic neuronal connections with learning, we increase the complexity of the system by creating additional ‘concepts’ or ideas, new options or associations.
Incorporating these four aspects into a real robot system is however a very demanding task, and not one that has yet been attempted. Firstly we cannot grow artificial parts, they do not ‘reproduce’ in any sense. The best we can do is either to ‘simulate’ such systems (perhaps eventually implementing the end result, e.g. Bentley, P.J., 2004), or to have our robot (somehow) pick up and incorporate extra parts that happen to be ‘lying around’ in the environment. Secondly it is unclear how we should have the parts interact in a way suitable for stimulating development, such that it can possibly start to self-organize itself. Thirdly we have the problem of how to allow the environment to change the configuration, to ‘disturb’ the robot in some way, in such a manner as to force it to re-design itself. Even if these three major obstacles are overcome, then we still do not know how to use our parts to self-assemble ‘critters’ with specific functions - but of course we still don’t know how nature does that either (Raff, R.A., 1996). This relates to understanding how the three processes (phylogeny, ontogeny and epigenesis) interact, but once we can do this then we have the potential to build what have been called ‘POEtic machines’ (Teuscher, C., 2001). For the future, perhaps all we can say, is that “we live in interesting times...”