Communication without dedicated signalling channels. A general finding?

The central finding of Quinn (2001) was that communication can evolve in an evolutionary robotics context without the use of dedicated signalling channels. Quinn simulated fairly realistic robotic agents controlled by neural networks and equipped with proximity sensors and wheels for locomotion.  The agents were set a coordinated movement task (i.e., to move their combined centre of mass as far as possible in a limited time frame).  Non-coordinated strategies do very poorly at this task, but coordination was not trivial to achieve, as the agents had no pre-given way of signalling to each other. Evolutionary runs revealed that coordinated overall behaviour could in fact emerge from a dance-like movement pattern that allowed the two agents to spontaneously establish “leader” and “follower” roles. Quinn’s result is very exciting because it shows the potential for ALife models to look at the origin of communication from genuinely non-communicative contexts. Other models that look at the conditions for the stability of a signalling system over a pre-defined signalling channel can only really refer to the evolutionary maintenance of communication rather than its beginnings. Although other evolutionary robotics researchers have referenced Quinn’s result, this has typically been in the context of interpreting some evolved behaviour in their own experiments. The importance of Quinn’s result for cognitive theorists interested in the evolution of language and social intelligence was acknowledged by Kirby et al. (2002) but this side of the work has not been pursued. Our project involves asking whether Quinn’s findings are general. In other words, we have successfully replicated Quinn’s central result but without using the particular simulation framework that he employed. Most of the results of our experiment match Quinn’s results and therefore it is suggested that the emergence of communication from non-communicative origins is likely to be a common evolutionary adaptation to niches that involve the coordination of cooperative behaviour.

DisertationFull

Emergence of Intelligence and Mind in an Artificial Living System

In 2000, a group of influential Artificial Life (AL) researchers compiled a list of problems, trying to reunite and synthesize the most important current open questions in AL (Bedau et al. 2000). Due to the high interdisciplinary nature of AL, many of the challenges proposed have a tight relationship with many different disciplines (computer science, mathematics, cognitive science, biology, ecology, chemistry, physics, etc.). Answering most of them would have important repercussions in many fields of knowledge. Among the fourteen questions proposed, the eleventh is particularly relevant due to its wide range of implications, including philosophical and ethical. The challenge states “Demonstrate the emergence of intelligence and mind in an artificial living system”. This is not a new challenge, it’s very old actually, and it’s shared between two main disciplines, AL and Artificial Intelligence (AI). When the AI research area was born in the 1950‘s, its motivations were around the same question and challenges: the building of machines with a general intelligence, hence being able capable to act like humans. As the AI field grew up, it became more and more divided into very specific subareas and, in many cases, with very little communication among them. At present, the AI research mainstream field, is much more concerned about engineering and optimization problems and its applications than with the initial problem. In the other hand current AL research interests feet very well with the challenge as well as Strong AI.

While the challenge is intuitively clear, the terms of mind and intelligence are not so. The question of mind, and by extension, consciousness, had been a recurrent topic in the history of human Though. The problem had been treated by many disciplines and by many different approaches among them. A definition of consciousness is not easy to achieve due to its intangible nature. David Chalmers divided the problem of consciousness in two different problems: the easy problem and the hard problem (Chalmers 1995). The easy problem is solvable and relates to the cognitive-computational aspects of the mind such image recognition, information integration, etc. The hard problem is related to why do we feel something associated with our perceptions of the world (Qualia). This clear dualism has been widely argued (Dennett et all 1997). Among them, Daniel Dennett denies the hard problem arguing the whole consciousness is the impact of experiences on behavior. A more pragmatic approach is given by Julian Jaynes (Jaynes 1976) and and more recently by Merlin Donald (Donald 2001), stating consciousness is in fact language, and no consciousness could exist without a high level of language complexity.In any case, consciousness and mind[1] are very difficult terms to define and measure, hence there is an inherent problem with this challenge: how to know if achieved or not.

Assuming complexity of living systems is a result of the adaptation to their environment, the development of minded individuals (and the mind itself), should be an adaptation result too. In the natural selection framework, the fittest individuals to a particular environment tend to prevail. Therefore, a particular environment would lead to a particular individual: it’s not probable to find eyed individuals in light absent environments. Following this reasoning line, a clear question emerges: what kind of problem (environment) did primitive humans (and other mammals) found, that drove them to develop a mind? Taking into account that very few species had developed something similar, the problem they faced should very uncommon. Not all environments produce the same amount of complexity as adaptation. In AL models, no complexity emerges applying therules of evolution by natural selection (at least, nothing comparable to natural complexity). The simulations tend to evolve and stay (forever?) in a stable configuration (Life, Terra, and others). These observations lead to two non-mutually exclusive conclusions. The first one, something is missing in evolution by natural selection theory (Bedau 2006). The second obvious conclusion is we are trying to reproduce results that took millions of years in a extremely complex environment with not enough resources.

Despite the large number of different technological and theoretical approaches that had been used to try to achieve a synthetic mind (Artificial Neural Networks (ANN’s), Recurrent Neural Networks (ARNN’s), Intelligent Agents, Rule Systems, etc.), we are going to discuss only two of them: models based on intelligent agents and a possible cellular automata model.

In 1994, Karl Sims developed a fully evolved realistic physical agent (Sims 1994). The aim of its work was to demonstrate that motor capacities could be evolved in a 3D realistic physical simulation. Using a Genetic Algorithm (GA) and a developmental representation, the individuals evolved on 3D physical simulation. In each generation, the population of creatures was selected maximizing the development of physical movement like walking, swimming and flying. Worst scored creatures of each generation were “killed” and the fittest were combined to give rise a new creatures. One of the key points in Sim’s model was how each of the individuals was represented. The genotypes of the creatures encoded both brain, in the form of a neural network, and body, making the whole creature highly evolvable: each of its components was susceptible to change in form of mutations and crossover.

Mind and language had probably very similar origins. The study of the origins of language and communication are fundamental to understand the origins of mind and maybe the mind itself (Jaynes 1976). In this sense, models that mimic the origins of language could be very useful to understand the evolution of consciousness and to study the role of language in the development of consciousness. From an evolutionary point of view, the origin of language is problematic because it contains a paradox: the existence of a signal has no sense if nobody can understand it (Maynard Smith 1997). No signal could exist without a response and no response could exist without a signal. Before 2001, all the models aimed to study the origins of language were designed with explicit communication channels. In 2001, Mark Quinn showed that communication between agents could be evolved without specific communication channels (Quinn 2001). In Quinn’s model, instead of using fully developmental representations of the individuals, only the brain, in form of a neural network, was evolved. The model was based in a population of simulated Khepera robots equipped with local sensors (short range IR). The individuals of the population are evaluated in pairs, maximizing their cooperation to solve a particular problem, so the fitness of each individual was highly coupled with its pair. The simulation showed that basic communication evolved from functional but non-communicative behavior. Another interesting observation is when communication evolved, always showed a hierarchy: one of the robots leads and the other follows, but none is intrinsically biased to adopt a role.

Intuitively, Cellular Automata (CA’s) do not seem to be an ideal framework for this goal, and at present, no cognitive model exists based in a strictly CA approach. CA’s are so simple that any sophisticated model using a CA would require enormous computing resources. Some specialized hardware architectures, like TM87, were designed improve the speed of the CA being simulated (Toffoli, Margolus 1987) but they were very far of the theoretical resources that big complex simulation would need (Poundstone 1985).

Imagine we’ve got an amazing complex simulation running in a very powerful cluster computing system, evolving with a state of the art genetic algorithm or even alone in a self-sustained universe. How can we know our model is conscious? The detection problem is clearly critical and it’s certainly non trivial. Since computational models are information based, a straightforward way of measuring its complexity would be through the Shannon’s Entropy. The problem with Shannon’s measure is that a totally random system would have maximum entropy, thus maximum complexity. In 1960’s, Andrei Kolmogorov proposed an algorithmic measure of complexity consisting on the measure of the size of the minimal program that can build the system being studied. Despite that Kolmogorov Complexity is uncomputable, it’s also maximized by randomness. Neither Shannon’s Entropy nor Kolmogorov Complexity measures capture structure or correlation from the system they are measuring. The intuitive relation between entropy and complexity has been proved false, no universal relationship between them exists (Feldman et al. 1998). Other measures based on similar approaches (algorithmically or thermodynamically) have been proposed but appear to be hardly computable or practical. The computational mechanics framework tries to combine the previous attempts into a new method to assess the complexity of a system. The Epsilon Machine (E-Machine) is the computational mechanics attempt to measure the effective complexity of a system (Crutchfield 1989). Another interesting complexity measures are based in the mutual information (a.k.a. transinformation) concept. In this context the complexity is a measure of the mutual dependence of two variables. A detailed study and evaluation of the different complexity measures is beyond the aim of this document, enough to say no complete measure exists yet to measure the sophistication in terms of information processing and hierarchy of a complex system.

Regardless of all, CA’s had fascinated researchers for the past 30 years because they show very interesting properties that make them ideal candidates for pure formal simulations. Due to they formal nature, a CA constitutes a universe by itself, with it’s own rules, universal computation capabilities (Berlekamp et al. 1982) and self-referencing abilities (Wolfram 2002). The entire research field of modern AL was founded around CA’s as the substrate for artificial life, starting with the Game of Life (Conway 1970) that was one of the first in-sillico successful models of artificial life.

An hypothetical large scale simulation based on a CA has potential to show all steps of evolution, from “bits” to “multi cellular organisms equivalent” and so on, providing an unified framework for modeling all the aspects of life (physical, biological and psychological) (Mandik 2008).

The evolution of a synthetic minded organism and therefore, the evolution of complexity comparable to biological systems is not a trivial task. Even with some assumptions regarding evolution issues, there are still too many critical open questions. The lack of a good measure of the complexity of a system is big drawback, it makes very difficult to accurate asses the model. The epistemological problems around the mind and consciousness concepts make difficult to have a well-defined target. In the other hand, the study of what kind of environments could lead to the evolution of mind would probably give valuable information. Language seems to have a strong paper in mind formation, their origins seem to be close (Jaynes 1976).

A hypothetical experiment is proposed. In the case we had a good method for measuring the complexity of a system (CA), what would happen if the system’s rules would be evolved through a GA in which the fitness function would maximize the amount of complexity? Even using the existing measures of complexity that take into account structure and hierarchy, the experiment would be interesting.

References

Bedau, M. A., Packard H., Rasmussen S., Adami C., Green, D. G., Ikegami T., Kaneko, K., Ray T.

S., “Open Problems in Artificial Life”, Artificial Life archive Volume 6, Issue 4, pp.363-376

Bedau, M. A. (2008). “The evolution of complexity”, Thomas Pradeu, et al, “Mapping the Future of Biology: Evolving Concepts and Theories”.

Berlekamp, E., Conway, J. H., and Guy, R. (1982): Winning Ways for Your Mathematical Plays, volume 2. Academic Press.

Chalmers, D. J. (1997), “The Conscious Mind “, Oxford University Press.

Crutchfield, J. P. (1990), "Reconstructing Language Hierarchies", in Information Dynamics, NATO ASI Series, 256, H. A. Atmanspracher and H. Scheingraber, editors, Plenum Press, New York, pp.45-60.

Crutchfield, J. P. and Young K. (1989), "Inferring Statistical Complexity", Physical Review Letters 63, pp.105-108.

Dennett, D., McGinn, C., Varela, F., Crick, F., and Penrose, R. (1997), “Explaining Consciousness: The Hard Problem.”, Shear, J. Explaining Consciousness: The Hard Problem. Cambridge: MIT Press.

Donald, M. (2001), ”A Mind So Rare he Evolution of Human Consciousness”, Norton, p.202.

Feldman, D. P., Crutchfield J. P. (1998), “Measures of statistical complexity: Why?”, Phys. Lett. A, 238:244-252.

Gardner, M. (1970), Issue of Scientific American, in Martin Gardner's "Mathematical Games"

Jaynes, J. (1976), “The Origin of Consciousness in the Breakdown of the Bicameral Mind”,  Mariner Books, p.55 ,p.65, p.66.

Langton, C. G. (1991), “Life at the Edge of Chaos”, Artificial Life II, SFI Studies in the Sciences of Complexity, vol. X, ed. C. G. Langton, C. Taylor, J. D. Farmer, S. Rasmussen, Addison-Wesley.

Maynard Smith, J. (1997), “The Theory of Evolution”, C.U.P., 3rd edition.

Mandik, P. (2008), “Cognitive Cellular Automata”, Complex Biological Systems: Applications in Real Life. Icfai University Press.

Poundstone, W. (1985), “The Recursive Universe”, William Morrow & Company, Inc.

Quinn M. (2001), "Evolving Communication without Dedicated Communication Channels"

Sims, K. (1994),  "Evolving Virtual Creatures", Siggraph '94 Proceedings, pp.15-22.

Toffoli, T., Margolus, N. (1987), “Cellular Automata Machines”, MIT Press.

Wolfram, S. (2002), “A new kind of science”, Wolfram Media, p.1192

---

[1] In this review, mind and consciousness are going to be used as a synonyms

The Decentralized Mindset

The archetypical example, ask someone: Who leads a flock of birds? (Resnick, 1994). You will have 99% chance to get the wrong answer. This is becoming a leitmotif on Complexity Science, and, by hence, on one way of thinking: the centralized mindset. It’s a simple question, then, why do everyone answers wrong? From children to adults, from officers to scientists. No matter who you ask, everyone will answer the same: “the first bird“. It’s been known for some years that no bird leads in a flock, the flock itself is a result of the interaction of the birds, it’s self organized (Potts, 1984). For someone belonging to the western culture, the correct answer is unintuitive. There’s a natural trend to look for an answer within the elements, while the next level of description is often omitted. In this case, the correct answer is related to the next level of description, the flock. In this text we are going to find the roots of such kind of thinking and study how it was maintained for such long period of time.

The centralized mindset has not always been the paradigm employed by the human being to explain his world and himself. More than 2500 years ago, even before the great philosophers Aristotle and Plato, the called pre-Socratic group, Heraclitus of Ephesus in particular, tried to find answers employing a systemic paradigm, focused in interrelationships and dynamic processes,  rather than a systematic one, centered in the concepts of classification and static order. In his doctrines of universal flux1 and the unity of opposites2, Heraclitus shows the unicity of opposites in a specific moment, time is the only factor of change.

Despite of pre-Socratics are the first sign of systemic though in the western culture, there are even more antiques examples. As far as 3200 years ago, the I Ching3, one of the chinese classic texts, is probably the oldest example of such kind of thinking. Its title itself “I Ching” means “Classic of Changes” and it’s a reference to its aim to model the dynamics between different elements as a whole4. The resemblances between Heraclitus thesis and I Ching principles show clearly that both cultures, in different moments of history, showed some degree of parallelism (Hammond, 2003).
This monistic understanding of the world can be found in most eastern religions-philosophies5. The collection is seen as a whole and not as the sum of its multiple parts. There is a natural systemic understanding of the world that flows from those philosophical frameworks.
Why two cultures starting from similar points, in terms of though, diverged so much? which was the cause of western culture became more and more dualistic, meanwhile eastern cultures stayed in a more holistic comprehension of the world?
One of the probables causes could be the incredible impact of the works of both Plato and Aristotle had in the western culture. Plato and Aristotle were more or less dualists, in the sense they argued to destroy the pre-Socratic monistic arguments. Unfortunately, in terms of historical success, no comparable figure to Plato and Aristotle existed who defended a systemic approach. That’s probably because all of the works of the pre-Socratics were lost, including Heraclitus6.
Plato is widely recognized to be the first to systemize the mind-body dualism in his Theory of Forms (Plato, Phaedo and other dialogs) in which he separated the perceived objects from their ideas or essences. Aristotle was contrary to the pre-Socratic monism. In Physics, Aristotle discusses against the natural pre-Socratic vision of change, in particular from Parmenides and Heraclitus, arguing that essences persisted through the change and opposites were different, separate thing (Aristotle, Physics, 184 b1).
After the classical period, the arrival and later imposition of the christian doctrine in Europe, constrained any philosophical position contrary to the christian catholic establishment. During this period, society was religion centered and most of the knowledge was controlled by the church, so philosophy were also god centered. In need of a more formal philosophical framework, christian philosophers belonging to the scholastic movement like Thomas Aquinas, incorporated Aristotelian ideas to the christian doctrine (McInerny, 2009), fixing a non-monistic approach in a religion that would rule Europe for 1000 years.
Starting in the seventeenth century, numerous attempts to re-introduce a systemic point of view were done by different scientists and philosophers. We can find monistic approached in authors like Spinoza, Berkley, Leibnitz and Hume. Despite most of them were very successful, in the seventeenth century the systematic approach was strongly rooted in Europe's society. The tremendous success of the systematic approach in science, achieved by Descartes7 and Newton8 among others, didn’t help.

It wasn’t until the twentieth century, when the limitations of the systematic approach became obvious. Scientists from all the areas, beginning from philosophy and social sciences, started to realize that a systemic approach, not only could help to understand, but maybe was the only method to gain knowledge in some disciplines. The centralized mindset, that had lot of success in the past, started to show it’s weaknesses: it was unable to give satisfying answers to important scientific problems9. The result of that need was the Systems Theory, an interdisciplinary framework whose main goal was to investigate all kind of disciplines from a systemic point of view.
In the first half of the twentieth century, the systemic approach was introduced in philosophy, sociology and economics by authors like Pareto, Spencer, Durkheim, Hartmann and others. Most of them explored the twentieth century advances, not from the classical Aristotelian-Platonic-Newtonian point of view, but from a systemic-monistic perspective, where the system gains relevance over the parts that is composed by. Around the 1950s, the tendency was extended to mathematics, physics and biology. Concepts like self-organization, complexity, connectionism, adaptive systems were explored by Wiener, Ashby, Neumann, Foerester, Lyapunov, Pointcare, and others. As the twentieth century advanced,  different flavors appeared: Cybernetics, Catastrophe Theory, Chaos Theory, Context Theory and Complexity Science. All of them were focused in studying what kind of patterns, behaviors and properties the systems show. They focused more and more on systems that showed complex behaviors, that’s behaviors that couldn’t be predicted observing the different elements of the system.

We’ve seen the origins of the centralized mindset are rooted in the more than 3000 years ago, when the western culture born. A bunch of circumstances and the big benefits it has reported to science helped to maintain it over more 2000 years.
The new born Complex Systems Science, seems to be the crystallization of the systemic approach.
Despite nowadays the systems approach is widely respected in many disciplines, the centralized mindset is still much more extended. Sometimes because it’s irrelevant and sometimes by ignorance. 2000 years of history of systematic thinking are not easy to fight with, and in the other side, it’s not an adequate framework to very important problems. Despite the systemic approach had been very successful in many disciplines, a breakthrough comparable to Einstein’s Relativity or Newton’s Laws hasn’t arrived yet. Over the twentieth century the systemic approach had been gaining critical mass, it’s just a matter of time that we will see an critical breakthrough, promoted by systems science.

References:
Graham, W. D. (2007), “Heraclitus, The Stanford Encyclopedia of Philosophy” (Winter 2003 Edition). Edward N. Zalta (ed.).
Hammond, D. (2003). “The Science of Synthesis”. University of Colorado Press (ed.). p. 23
Kahn, C (1979). “The Art and Thought of Heraclitus: Fragments with Translation and Commentary”. Cambridge University Press (ed.). pp. 1–23.
McInerny R. (2009), “Saint Thomas Aquinas, The Stanford Encyclopedia of Philosophy” (Winter 2003 Edition). Edward N. Zalta (ed.).
McKeon. R. (1941). “Basic Works of Aristotle”. Random House (ed.). pp. 34-56
Plato (390s-347 BC) Platonis Opera, vol. 1, Euthyphro, Apologia Socratis, Crito, Phaedo, Cratylus, Theaetetus, Sophistes, Politicus, ed. E.A. Duke, W.F. Hicken, W.S.M. Nicoll, D.B. Robinson and J.C.G. Strachan. Clarendon Press (ed.), 1995.
Potts, W. K. (1984). “The chorus-line hypothesis of coordination in avian flocks”. Nature 24: 344-345.

Resnick, M. (1994). “Turtles, Termites, and Traffic Jams: Explorations in Massively Parallel Micro worlds”. MITPress (ed.).