I present a qualitative evolutionary model based on a new theory of natural causation. The model promises a solution to a notorious complexity bottleneck in simulated evolution. I discuss some implications for evolutionary technology.

This paper is concerned with fundamentals of evolutionary theory with an eye to applications in modeling and technology. Evolution is considered here not in the limited sense as genetic transformation but in the more general sense as the appearance of new life forms or ‘species’. From the engineering point of view, a species is a family of material functions that it can perform. Evolutionary simulations of the type exemplified by genetic algorithms and other selectionist systems tend to be non-productive in terms of the emergence of new material functions. These simulations ‘bottom out’ very soon at some (possibly optimal) solution of the given particular functions, or enter a stationary domain of complicated behaviors characterized by unstable solutions of the same functions while competing for a given resource (such as space). Evolutionary models based exclusively on artificial selection apparently cannot overcome this problem. A solution, based on natural causation, is presented in this paper. I suggest that in biological systems it is natural causation that ensures an unlimited supply of new selection forces that lead to new material functions. Natural causation is ‘deep’causation, which means that every causal process is manifested as a unity of several different cause-and-effect relations. This allows for a model of evolution based on overlapping sub-evolutionary selection processes perpetuated by the causal nature of the evolutionary transitions.

Evolutionary biology most often speaks of genetic transformation in a population. Yet in a broader sense it is clear that evolution also involves something else, as already obvious from the title words of Darwin’s classic work The Origin of Species. Along a different thread, another key problem of modern evolutionary theory is the origin of evolutionary complexity (Maynard Smith 1982, [1]). These two aspects, the problems of new species and more complexity, are linked together by the recognition that a species is a set of material functions that it can perform. Thus, the origin of species and the origin of evolutionary complexity have a common root: the emergence of new material functions in the lineage of organisms. Historically, this idea is captured in the notion of ‘design’ (see Dennett 1995, [2]), which also helps to better clarify the concept of material function. A material function is something that can fulfill a purpose.

The origin and the development of design, or functional complexity, is what evolutionary technology, in particular, could utilize. It is tempting to imagine, for instance, a set of evolving ‘seeds’ dispatched over a territory, so that in the course of time new and more complicated life forms can emerge that perform complex tasks. Perhaps such a process can be controlled or predicted; in such a way the potential ‘amplification’ effect of evolution (that one gets more material functions than one has ‘paid for’) can be combined with an ‘explanatory transparency’ required by engineering.

The most widespread form of technological amplification in evolution is the development of new organs and the functional differentiation of the organism as well as the population to which the organism belongs. Natural evolution provides a clear example for such a process. Can we repeat or simulate this process?

In the modern synthetic theory evolution is understood as selective survival based on a reproductive differential associated with heritable traits. The engineering notion that this concept has become is that of adaptive search. Under this hypothesis, an evolutionary process is convergence towards a goal or a set of complicated goals (as in the case of rugged fitness functions and elsewhere). It is also understood that goals themselves can be changed, such as in co-evolutionary models (Van Valen 1973, [3]). Yet the behavior of simulated systems based on the current theory is frustrating when compared to the broader vision of evolution. Adaptation produces fast results but stops afterwards, in lack of a new input from selection; new goals, insofar as they exist at all, tend to be variations of the old ones. Together, these imply a failure to catch the essence of evolution.

It is easy to find examples to illustrate this point. Most known realizations of evolutionary principles, such as genetic algorithms, Dawkins’ ‘Weasel’ systems [4], Tierra-like computer memory competitions [5,6], or Pollack’s Golem project [7] all suffer from the same problem. Let us reivew two of them briefly. Tierra started out as a major promise for artificial evolution and it generated several interesting outcomes (e.g. the emergence of arms race and parasitism in a population) but it soon became clear that the system fails to produce a sustained evolution process, as is evident to anyone who has ever experimented with it. Running the system for days (or on hundreds of computers simultaneously) usually makes things worse, as was the case with the Golem project, our second example.

The Golem project, an initiative now abandoned, was a recent experiment in synthetic evolution with the aim to demonstrate the possibility of robotic development using evolutionary programming. Initially, the system produced various ingenious solutions to the selected problem, locomotion, but despite various efforts it failed to improve any further. Pollack concluded that “merely more CPU is not sufficient to evolve complexity” (Pollack 2002 [8])).

In other words, evolutionary simulations tend to deplete their potential, or indeed their resources for development. There is something like a ‘Zeno paradox’ of these simulations: the evolving population can never pass a complexity barrier because each selection step is measured against that same barrier. Can we think of pulling a toy train with a magnet, and pushing the same magnet by the train, simultaneously? It is impossible to do both at the same time. One needs an independent source of fuel, but where does the fuel come from? Apparently, to answer this question for evolution, one has to extend the adaptation and optimization paradigm.

Natural evolution is a causal process, and an analysis of natural causation may help us to understand and overcome the functional complexity bottleneck of selectionist systems.

Understanding causation is of key importance for science in general. A detailed treatment of the problem is not the target of this paper. A brief and simplified characterization of causality will be sufficient for our purpose. A most important point is that natural causation is not just a relation between a cause and an effect, or between a state and a future state. For instance, an equation can generate new values yet equations are not causal (as recognized by people as different as B. Russell or more recently J. Pearl 2000, [9]). A well-known way to express the difference is to say that causality pertains to manipulation and experiment, and not to description (Hacking 1983, [10]). A recent study (Kampis 2002 [11,12,13,14]) identified some key features of causation that operationalize this intuition, with a relevance to our present topic. The central idea is that of ‘depth’.

Causal processes in Nature are always multiple or ‘deep’. By this I mean that every causal process does several things at once. This feature of causality can be expressed as a modal unity of (indefinitely many) simultaneous cause-effect relations. A simple case here is that of ‘object-causation’, the situation where a given causal relation among events depends on some object on which the events are defined. This automatically generates further, parallel, causal relations that involve further events. The further causal relations are generated by virtue of the property that the involved events all belong to the same object.

It is important to note that the point is not the existence of continuing causal chains, or other future consequences of the given causal relation, but rather the simultaneity of several causal subprocesses within it.

Here is a concrete example, taken from everyday life. It’s straightforward to generalize for science. If I violently push my armchair against the desk, this action gives rise to the causal pair of events that the displacement of the chair is followed by that of the desk. But my action also gives rise to the causal pair of events that the motion of the chair’s arm moves the color patches on top of the desk, left by the rim of my coffe cup. In the same action, magnetic events in the chair can cause electric events in the desk, which is particularly easy to see if the latter is a metallic bureau piece. And so on. There are always countlessly many different kinds of events that can be defined over every causal transition, and they generate equally numerous different cause-effect pairs. These together constitute what we recognize as one causal process.

Also the opposite is true, a relation between events will never termed causal unless it is accompanied with ‘multiplicity’or ‘depth’. In particular, it is easy to show by means of counterexamples that counterfactual dependence is neither sufficient nor necessary for causality (for a first presentation, see Kampis 2002 [14]).

By contrast to this, artificial selectionist models are clearly non-causal in the sense that they exemplify an attempt to cope with evolution in terms of single-aspect causation, which is nowhere found in Nature.

It will be instructive to combine the understanding of causation with evolution. First let me comment on the nature of natural selection to avoid confusion. Selection has always been just a shorthand for the effect of a ‘dominant’ selection force, singled out from a more complicated set of interactions. In every evolutionary situation there are numerous selection forces, but in a model we assume that others than the dominant are slower, less dramatic, closer to local optima, or can be ignored on some other basis. I will keep this simplification. With the aid of the tool of natural causality we will obtain a picture where dominant selection forces can change. However, this is not due to a change of our attention. It follows from the properties of causal interaction.

Usually, it is assumed that selection processes only operate on the traits to which they are applied. However, it is clear that natural selection takes the form of a causal transformation. As a consequence, it produces a multiplicity of events. That is, besides leading to adaptation in certain traits, the process simultaneously transforms (and perpetuates) a set of other traits, to which currently no selection is applied. In other words, instead of a single-aspect process, as defined by the selective force alone, we deal with a ‘deep’ process here, defined by the total causal relation that realizes the selection force. This will make it possible for evolution to feed back the products of selection to the very selection process.

When changes are classified as adaptive and non-adaptive, this refers to a given point of time, and to a given selective force. Non-adaptive changes brought forth by natural causation can act as a reserve for future development. When one domain of selection is depleted, another can take over, using a new selective force defined by traits modified as a consequence of causal depth.

In other words, natural selection produces not only adaptations but also new causal opportunities, and as the selection process continues, the accummulating causal effects can provide a new bottleneck of survival, so the process may start over again. Using suggestions from this picture, it is now possible to outline a new, causal model of evolution.

.

Non-adaptive changes in the non-selective traits provide a possibility for sustained evolution in one of the following ways.

1. Changing non-selective traits are open for new environmental interaction that influences survival.
2. Changing non-selective traits provide changing environment and new selection force for other organisms.

We can combine these into a single causal evolutionary mechanism:

(3) Change in non-selective traits induces change in selective traits.

This combined formulation leaves open the important question of where do the new evolutionary effects occur. These can be effects on the given individual species, or effects on some other species, such as in co-evolution. (In a simulation context, these would correspond to effects on ‘organisms’ of the same type, or ‘organisms’ of some different type, respectively.) Whether a sustained evolution process is possible at the level of one individual species (one ‘organism type’) is a further topic to study, and not the target of this exploratory paper.

We can summarize the obtained insights.

It is a long way to go from the still-dominating fully abstract systems to the required causal systems, but some first suggestions can already be articulated.

In evolution, a synonym for causal interaction is phenotypic interaction. However, evolution is a genetic process. This genetic process is made possible by the existence of what is called a phenotype-genotype mapping. From our previous discussion we can safely conclude that the phenotye-genotype mapping itself has to be causal. Otherwise, it would be difficult to make sure that any causal effect (in the sense of ‘depth’) that influences the phenotype could directly propagate to the genotype. Such a propagated influence is a precondition for the start of any new selection process.

Life as we know it solved this problem by equipping organisms with a body that is actively maintained. Therefore, any causal interaction that affects the phenotype can potentially interfere with the genes by affecting survival through affecting self-maintenance – in evolution theory, this is usually called ‘hard selection’ (i.e. selection that influences the chance of endurance and the chance of reproduction of the organism). If the causal model is correct, it suggests that conventionally built simulations (and conventionally built robots) may not be suitable for evolutionary experiments just by having self-reproduction and mutation. These systems currently lack a feedback from the simulated evolution process to their existence, which can now be suspected as a requirement for causal evolution to proceed. Several suggestions exist about the relevance of self-maintenance (that is, metabolism) in life – the present discussion suggests a future convergence of this topic with causal evolution, an issue that requires further study.

The qualitative causal model can be conceived as an extension of the well known co-evolutionary theories where adaptive gain in one species is experienced as a fitness loss in other species. There is a significant difference, however. Here causality offers a mechanism for qualitative novelty. Current simulations, or perhaps any simulations of whatever kind, lack this kind of property, since they lack causality – that is, they lack a unity offered for free by Nature as part of its material constitution.

Perthaps this is what matters: in short, matter matters. If that is true, evolutionary simulations and future evolutionary technologies have to take a broad leap, similar to the one robotics had taken some time ago, and for a similar reason. Situated robots (in the sense of R. Brooks and others [15]) were the first to possess the ability for non-prescribed performances, which turned out to be critical for the understanding – and the exploration – of what robots can do. The same could be expected in evolution when replacing artificial selection with the complexity of the natural causal process.

The author wishes to thank the Fujitsu Company and Professor Susumu Kunifuji for the support that made this paper possible. Special thanks to Professor Ichiro Tsuda for his helpful comments and suggestions.

[1] Maynard Smith, J. (ed.) 1982: Evolution Now, Macmillan, New York.

[2] Dennett, D.C. (1995): Darwin’s Dangerous Idea, New York.

[3]Van Valen, L. 1983: A New Evolutionary Law, Evolutionary Theory 1, 1-30.

[5] Ray, T. S. (1992): Evolution, ecology and optimization of digital organisms, Santa Fe Institute working paper 92-08-042.

[6] Ray, T. S. & Hart, J (1998): Evolution of Differentiated Multi-threaded Digital Organisms. In: Artificial Life VI proceedings, C. Adami, R. K. Belew, H. Kitano, and C. E. Taylor (eds.), 295-304, MIT Press, Cambridge.

[7] Lipson, H. and Pollack, J. B. (2000): "Automatic design and Manufacture of Robotic Lifeforms", Nature 406, pp. 974-978.

[8] Pollack, J. (2002): sentence quoted from

[9] Pearl, J. (2000): Causality. Models, Reasoning, and Inference, Cambridge University Press.

[10] Hacking, I. (1983): Representing and Intervening, Caambridge University Press.

[12] Kampis, G. (2002): “Representation, Causality and Complexity”, JAIST lecture, 2002.04.25.,

[15] Brooks, R. A. (1991): Intelligence Without Representation, Artificial Intelligence Journal (47), pp. 139–159.