The Artificial Life Route to Artificial Intelligence

Building Embodied, Situated Agents

Luc Steels, Rodney Brooks, Luc Steels, Rodney Brooks

• Publisher: Routledge
• Year: 2018
• Print ISBN:
  9781138545786
• eBook ISBN:
  9781351001861
• Topic: Psychology
• Subtopic: Cognitive Psychology & Cognition
• Index: Psychology

Originally published in 1995, this volume is the direct result of a conference in which a number of leading researchers from the fields of artificial intelligence and biology gathered to examine whether there was any ground to assume that a new AI paradigm was forming itself and what the essential ingredients of this new paradigm were. A great deal of scepsis is justified when researchers, particularly in the cognitive sciences, talk about a new paradigm. Shifts in paradigm mean not only new ideas but also shifts in what constitutes good problems, what counts as a result, the experimental practice to validate results, and the technological tools needed to do research. Due to the complexity of the subject matter, paradigms abound in the cognitive sciences -- connectionism being the most prominent newcomer in the mid-1980s.

This workshop group was brought together in order to clarify the common ground, see what had been achieved so far, and examine in which way the research could move further. This volume is a reflection of this important meeting. It contains contributions which were distributed before the workshop but then substantially broadened and revised to reflect the workshop discussions and more recent technical work. Written in polemic form, sometimes criticizing the work done thus far within the new paradigm, this collection includes research program descriptions, technical contributions, and position papers.

1. The Re-Enchantment of the Concrete

Some Biological Ingredients for a Nouvelle Cognitive Science

FRANCISCO J. VARELA
Ecole Polytechnique

1.1 Shifts in Cognitive Science

Rationalistic, Cartesian, or objectivist: These are some terms used to characterize the dominant tradition within which we have grown in recent times. Yet when it comes to a re-understanding of knowledge and cognition I find that the best expression to use for our tradition is abstract: Nothing characterizes better the units of knowledge that are deemed most natural. It is this tendency to find our way toward the rarified atmosphere of the general and the formal, the logical and the well defined, the represented and the planned-ahead, that makes our Western world so distinctly familiar.

The main thesis I pursue here is that there are strong indications that the loose federation of sciences dealing with knowledge and cognition—the cognitive sciences—are slowly growing in the conviction that this picture is upside down and that a radical paradigmatic or epistemic shift is rapidly developing. At the very center of this emerging view is that the proper units of knowledge are primarily concrete, embodied, lived. This uniqueness of knowledge, its historicity and context, is not a noise that occludes the brighter pattern to be captured in its true essence, an abstraction. The concrete is not a step toward anything: It is how we arrive and where we stay.

Let me unfold this emerging view, which revitalizes the role of the concrete by focusing on its proper scale: the cognitive activity as it happens in a very special space that we may call the hinges of the immediate present, for it is in the immediate present that the concrete actually lives. But before this unfolding we need to revise some entrenched assumptions inherited from the computationalist orthodoxy.

1.2 Minds and Disunited Subjects

If we turn to consider the living, there is considerable support for the view that brains are not logical machines, but highly cooperative, unhomogeneous, and distributed networks. The entire system resembles a patchwork of subnetworks assembled by a complicated history of tinkering, rather than an optimized system that results from some clean unified design. This kind of architecture also suggests that instead of looking for grand unified models for all network behaviors, one should study networks whose abilities are restricted to specific, concrete cognitive activities that interact with each other.

This view of cognitive architecture has begun to be taken seriously by cognitive scientists in various ways. For example, as is well known Minsky [15] presented a view in which minds consist of many agents whose abilities are quite circumscribed: Each agent taken individually operates only in small-scale or toy problems. The problems must be of a small scale because they become unmanageable for a single network when they are scaled up. This last point has not been obvious to cognitive scientists for long time. The task, then, is to organize the agents, who operate in these specific domains, into effective larger systems or agencies and then to turn these agencies into higher level systems. In doing so, mind emerges as a kind of society.

It is important to remember here that, although inspired by a fresh look at the brain, this is a model of the mind. In other words, it is not a model of neural networks or societies; it is a model of the cognitive architecture that abstracts (again!) from neurological detail and hence from the wet of the living and of lived experience. Agents and agencies are not, therefore, entities or material processes; they are abstract processes or functions. The point bears emphasizing, especially because Minsky sometimes wrote as if he was talking about cognition at the level of the brain. As I emphasize, what is missing is the detailed link between such agents and the incarnated coupling, by sensing and acting, that is essential to living cognition. But let us pause for the moment to follow some of the implications of the notions of fragmented and local cognitive subnetworks.

The model of the mind as a society of numerous agents is intended to encompass a multiplicity of approaches to the study of cognition, ranging from distributed, self-organizing networks up to the classical, cognitivist conception of symbolic processing. This encompassing view challenges a centralized or unified model of the mind, whether in the form of distributed networks, at one extreme, or symbolic processes, at the other extreme. This move is apparent for example when Minsky argued that there are virtues not only in distribution, but in insulation, (i.e., in mechanisms that keep various processes apart¹). The agents within an agency may be connected in the form of a distributed network, but if the agencies were themselves connected in the same way they would, in effect, constitute one large network whose functions were uniformly distributed. Such uniformity, however, would restrict the ability to combine the operations of individual agencies in a productive way. The more distributed these operations are, the harder it is to have many of them active at the same time without interfering with each other. These problems do not arise, however, if there are mechanisms to keep various agencies insulated from each other. These agencies would still interact, but through more limited connections.

The details of such a programmatic view are, of course, debatable. But the overall picture it suggests is that of mind not as a unified, homogenous entity, or even as a collection of entities, but rather as a disunified, heterogenous collection of processes. Elsewhere I have discussed in extenso some important consequences of this idea [20]. Such a disunified assembly can obviously be considered at more than one level. What counts as an agency, (i.e., as a collection of agents) could, if we change our focus, be considered as merely one agent in a larger agency. And conversely, what counts as an agent could, if we resolve our focus in greater detail, be seen to be an agency made up of many agents. In the same way, what counts as a society will also depend on our chosen level of focus.

Having thus set the stage for this key issue in contemporary cognitive science, I want to develop its implications for the question at hand: the present-centeredness of the concrete.

1.3 Readiness-to-Action in the Present

My present concern is with one of the many consequences of this view of the disunity of the subject, understood as a cognitive agent. The question I have in mind can be formulated thus: Given that there is a myriad of contending subprocesses in every cognitive act, how are we to understand the moment of negotiation and emergence when one of them takes the lead and constitutes a definite behavior? In more evocative terms, how are we to understand the very moment of being there when something concrete and specific shows up?

Picture yourself walking down the street, perhaps going to meet somebody. It is the end of the day, and there is nothing very special in your mind. You are in a relaxed mood, in what we may call the readiness of the walker who is simply strolling. You put your hand into your pocket, and suddenly you don’t find your wallet where it usually is. Breakdown: You stop, your mind setting is unclear, your emotional tonality shifts. Before you know it, a new world emerges: You see clearly that you left your wallet in the store where you just bought cigarettes. Your mood shifts now to one of concern for losing documents and money; your readiness-to-action is now to go back to the store quickly. There is little attention to the the surrounding trees and passersby; all attention is directed to avoiding further delays.

Situations like this are the very stuff of our lives. We always operate in some kind of immediacy of a given situation: Our lived world is so ready-at-hand that we don’t have any deliberateness about what is and how we inhabit it. When we sit at the table to eat with a relative or friend, the entire complex knowhow of handling table utensils, the body postures, and pauses in the conversation, are all present without deliberation. Our having-lunch-self is transparent.² You finish lunch, return to the office, and enter into a new readiness with a different mode of speaking, postural tone, and assessments. We have a readiness-to-action that is proper to every specific lived situation. New modes of behaving and the transitions or punctuations between them correspond to mini- (or macro-) breakdowns we experience constantly.

I refer to any such readiness for action as microidentities and their corresponding microworlds. Thus, the way we show up as is the way things and others show up to us. We could go through some elementary phenomenology and identify some typical microworlds within which we move during a normal day. The point is not to catalogue them but rather to notice their recurrence: Being capable of appropriate action is, in some important sense, a way in which we embody a stream of recurrent microworld transitions. I am not saying that there aren’t situations where recurrence does not apply. For example when we arrive for the first time in a foreign country there is an enormous lack of readiness-to-hand and recurrent microworlds. Many simple actions such as social talk or eating have to be done deliberately and learned. In other words, microworlds and identities are historically constituted. But the pervasive mode of living consists of the already constituted microworlds that compose our identities. Clearly there is a lot more that should be explored and said about the phenomenology of ordinary experience.³

My intention here is more modest, merely to point to a realm of phenomena, that is intimately close to our ordinary experience: When we leave the realm of our lived human experience and shift our focus to animals the same kind of analysis applies as an external account. The extreme case is illustrative: Biologists have known for some time that invertebrates have a rather small repertoire of behavior patterns. For example, the locomotion of a cockroach has only a few fundamental modes: standing, slow walking, fast walking, and running. Nevertheless this basic behavioral repertoire makes it possible for these animals to navigate appropriately in any possible environment known on the planet, natural or artificial. The question for the biologist is then: How does the animal decide which motor action to take in a given circumstance? How does its behavioral selection operate so that the action is appropriate? How does the animal have the common sense to assess a given situation and interpret it as requiring running as opposed to slow walking?

In the two extreme cases, human experience during breakdowns, and animal behaviors at moments of behavioral transitions, we are confronted—in vastly different manners to be sure—with a common issue: At each such breakdown, the manner in which the cognitive agent will next be constituted is neither externally decided nor simply planned ahead. It is a matter of commonsensical emergence, of autonomous configurations of an appropriate stance. Once a behavioral stance is selected or a microworld is brought forth, we can more clearly analyze its mode of operation and its optimal strategy. In fact, the key to autonomy is that a living system, out of its own resources, finds its way into the next moment by acting appropriately. The breakdowns, the hinges that articulate microworlds, are the source of the autonomous and creative side of living cognition. Such common sense, then, needs to be examined at a microscale, at the moments where it actualizes during breakdowns, the birthplace of the concrete. This is, to be sure, also a central question for the design of autonomous robots [14], and it will be interesting to see to what extent similar solutions might not apply.

1.4 Knowledge as Enaction

Let me explain what I mean by the word embodied, highlighting two main points: (a) that cognition depends on the kinds of experience that come from having a body with various sensorimotor capacities; and (b) that these individual sensorimotor capacities are themselves embedded in a more encompassing biological and cultural context. These two points were already introduced when discussing breakdown and common sense, but here I explore further their corporeal specificity, to emphasize once again that sensory and motor processes, perception and action, are fundamentally inseparable in lived cognition, and not merely contingently linked in individuals.

In order to make my ideas more precise, let me give a preliminary formulation of what I mean by an enactive approach to cognition [19, 20]. In a nutshell, the enactive approach consists of two key points: (a) that perception consists in perceptually guided action; and (b) that cognitive structures emerge from the recurrent sensorimotor patterns that enable action to be perceptually guided. These two statements will become transparent as we proceed.

Let us begin with the notion of perceptually guided action. For the dominant computationalist tradition, the point of departure for understanding perception is typically abstract: the information processing problem of recovering predetermined properties of the world. In contrast, the point of departure for the enactive approach is the study of how the perceiver can guide its actions in its local situation. Because these local situations constantly change as a result of the perceiver’s activity, the reference point for understanding perception is no longer a predetermined, perceiver-independent world, but rather the sensorimotor structure of the cognitive agent, the way in which the nervous system links sensory and motor surfaces. It is this structure—the manner in which the perceiver is embodied—rather than some predetermined world, that determines how the perceiver can act ...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Table of Contents
• 1 The Re-Enchantment of the Concrete
• I Research Programmes
• II Technical Contributions
• III Position Papers
• Index