Randall Beer

Studies of mechanisms of adaptive behavior generally focus on neurons and circuits. But adaptive behavior also depends on interactions among the nervous system, body and environment: sensory preprocessing and motor post-processing filter inputs to and outputs from the nervous system; co-evolution and co-development of nervous system and periphery create matching and complementarity between them; body structure creates constraints and opportunities for neural control; and continuous feedback between nervous system, body and environment are essential for normal behavior. This broader view of adaptive behavior has been a major underpinning of ecological psychology and has influenced behavior-based robotics. Computational neuroethology, which jointly models neural control and periphery of animals, is a promising methodology for understanding adaptive behavior.

Using the language of dynamical systems theory, a general theoretical framework for the synthesis and analysis of autonomous agents is sketched. In this framework, an agent and its environment are modeled as two coupled dynamical systems whose mutual interaction is in general jointly responsible for the agent's behavior. In addition, the adaptive fit between an agent and its environment is characterized in terms of the satisfaction of a given constraint on the trajectories of the coupled agent-environment system. The utility of this framework is demonstrated by using it to first synthesize and then analyze a walking behavior for a legged agent.

Dynamical ideas are beginning to have a major impact on cognitive science, from foundational debates to daily practice. In this article, I review three contrasting examples of work in this area that address the lexical and grammatical structure of language, Piaget’s classic ‘A-not-B’ error, and active categorical perception in an embodied, situated agent. From these three examples, I then attempt to articulate the major differences between dynamical approaches and more traditional symbolic and connectionist approaches. Although the three models reviewed here vary considerably in their details, they share a focus on the unfolding trajectory of a system’s state and the internal and external forces that shape this trajectory, rather than the representational content of its constituent states or the underlying physical mechanisms that instantiate the dynamics. In some work, this dynamical viewpoint is augmented with a …

The" intelligence" of traditional artificial intelligence systems is notoriously narrow and inflexible--incapable of adapting to the constantly changing circumstances of the real world. Although traditional artificial intelligence systems can be successful in narrowly prescribed domains, they are inappropriate for dynamic, complex domains, such as autonomous robot navigation.** This book proposes an alternative methodology for designing intelligent systems based on a model of intelligence as adaptive behavior. The author describes an experiment in computational neuroethology--the computer modeling of neuronal control of behavior--in which the nervous system for an artificial insect is modeled. The experiment demonstrates that simple, complete intelligent agents are able to cope with complex, dynamic environments--suggesting that adaptive models of intelligence, based on biological bases of adaptive behavior, may prove to be very useful in the design of intelligent, autonomous systems. Provides a lucid critique of traditional artificial intelligence research programs Presents new methodology for the construction autonomous agents, which has implications for mobile robotics Of interest to researchers in a variety of fields: artificial intelligence, neural networks, robotics, cognitive science, and neuroscience

We would like the behavior of the artificial agents that we construct to be as well-adapted to their environments as natural animals are to theirs. Unfortunately, designing controllers with these properties is a very difficult task. In this article, we demonstrate that continuous-time recurrent neural networks are a viable mechanism for adaptive agent control and that the genetic algorithm can be used to evolve effective neural controllers. A significant advantage of this approach is that one need specify only a measure of an agent's overall performance rather than the precise motor output trajectories by which it is achieved. By manipulating the performance evaluation, one can place selective pressure on the development of controllers with desired properties. Several novel controllers have been evolved, including a chemotaxis controller that switches between different strategies depending on environmental conditions, and …

Notions of embodiment, situatedness, and dynamics are increasingly being debated in cognitive sci ence. However, these debates are often carried out in the absence of concrete examples. In order to build intuition, this paper explores a model agent to illustrate how the perspective and tools of dynam ical systems theory can be applied to the analysis of situated, embodied agents capable of minimally cognitive behavior. Specifically, we study a model agent whose “nervous system” was evolved using a genetic algorithm to catch circular objects and to avoid diamond-shaped ones. After characterizing the performance, behavioral strategy and psychophysics of the best-evolved agent, its dynamics are analyzed in some detail at three different levels: (1) the entire coupled brain/body/environment sys tem; (2) the interaction between …

Dynamical neural networks are being increasingly employed in a variety of contexts, including as simple model nervous systems for autonomous agents. For this reason, there is a growing need for a comprehensive understanding of their dynamical properties. Using a combination of elementary analysis and numerical studies, this article begins a systematic examination of the dynamics of continuous-time recurrent neural networks. Specifically, a fairly complete description of the possible dynamical behavior and bifurcations of one- and two-neuron circuits is given, along with a few specific results for larger networks. This analysis provides both qualitative insight and, in many cases, quantitative formulas for predicting the dynamical behavior of particular circuits and how that behavior changes as network parameters are varied. These results demonstrate that even small circuits are capable of a rich variety of …

Current debates regarding the possible cognitive impli-cations of ideas from adaptive behavior research and dynamical systems theory would benefit greatly from a careful study of simple model agents that exhibit minimally cognitive behavior. This paper sketches one such agent, and presents the results of preliminary experiments on the evolution of dynamical neural networks for visually-guided orientation, object discrimination and accurate pointing with a simple manipulator to ob-jects appearing in its field of view.

WHEN IT COMES TO AUTONOMOUS ROBOTS, THE CONTRAST between fantasy and reality is really quite striking. On the one hand, we have the fantasy of such anthropomorphic robots as C3PO from Star Wars and Commander Data from Star Trek: The Next Generation which, despite their endearing quirks, negotiate complex physical and social environments with essentially the skill of a human being. On the other hand, we have the reality of industrial robots that can efficiently carry out such highly specialized tasks as painting and welding only in environments carefully constrained to minimize complications. Or, to consider a second example, we have Dante II, the semi-autonomous robot that had to be lifted out of a volcano it was exploring with a crane when it overturned. Far from being

86 June 1999/Vol. 42, No. 6 COMMUNICATIONS OF THE ACM tive interactions with the people responsible for these other components. For example, implementing embedded control software for an automobile’s fuel injection system is very different from implementing a depth-first traversal of a binary tree in a data structures class. More generally, engineering and science students typically find the transition from student to professional difficult for four major reasons:(1) students are not trained to deal with problems that require an integrated approach;(2) they are rarely exposed to the real-world issues that such problems pose;(3) they are rarely encouraged to solve problems through teamwork, or to bring to bear information from multiple disciplines; and (4) they rarely have an opportunity for critical thinking.Integrated approaches are essential for solving problems in engineering and science. A tacit assumption of many engineering students is that if each piece of a complex project works in isolation, the complete system will work as a unified whole. This is almost never true in practice. Unexpected problems emerge unless one takes into account the special properties of each piece of the system, and their interactions. There is a growing need to deal with such problems. Increasingly, software is part of an embedded system such as an airplane, a microwave oven, or a copy machine. In such systems, it is essential to integrate (and trade off between) the design and implementation of mechanics, electronics, and control. Similarly, the education of biology students emphasizes the reductive analysis of animals into cells and molecules, but not how such …

Distributed control and local leg reflexes enable insects to cope easily with terrain that would defeat many legged robots. An insect-like hexapod robot incorporating biologically based control effectively responded to mechanical perturbations using active and passive compliance and a local stepping reflex. An elevator reflex and a searching reflex addressed unexpected obstacles and loss of support, respectively. The robot exhibited a range of gaits using stick-insect-based distributed control mechanisms and negotiated irregular, slatted and compliant surfaces with this biologically based control strategy.

Expandable actuators surround a central conduit. Each actuator comprises a bladder that, when fluid is introduced, expands laterally while contracting longitudinally. A restorative spring can be placed inside a bladder and between the two ends to restore the actuator to its original shape as fluid is withdrawn. Multiple actuators can be placed in series to successively inflate and deflate and generate a peristaltic motion. One or more Shape Memory Alloy (SMA) springs can be affixed to one or more restorative springs to cause bending motion.

This article describes techniques that have been developed for spatial learning in dynamic environments and a mobile robot system, ELDEN, that integrates these techniques for exploration and navigation. In this research, we introduce the concept of adaptive place networks, incrementally-constructed spatial representations that incorporate variable-confidence links to model uncertainty about topological adjacency. These networks guide the robot's navigation while constantly adapting to any topological changes that are encountered. ELDEN integrates these networks with a reactive controller that is robust to transient changes in the environment and a relocalization system that uses evidence grids to recalibrate dead reckoning.

Of the various attempts to generalize information theory to multiple variables, the most widely utilized, interaction information, suffers from the problem that it is sometimes negative. Here we reconsider from first principles the general structure of the information that a set of sources provides about a given variable. We begin with a new definition of redundancy as the minimum information that any source provides about each possible outcome of the variable, averaged over all possible outcomes. We then show how this measure of redundancy induces a lattice over sets of sources that clarifies the general structure of multivariate information. Finally, we use this redundancy lattice to propose a definition of partial information atoms that exhaustively decompose the Shannon information in a multivariate system in terms of the redundancy between synergies of subsets of the sources. Unlike interaction information, the atoms of our partial information decomposition are never negative and always support a clear interpretation as informational quantities. Our analysis also demonstrates how the negativity of interaction information can be explained by its confounding of redundancy and synergy.

We are interested in the synthesis of autonomous agents using evolutionary techniques. Most work in this area utilizes a direct mapping from genotypic space to phenotypic space. In order to address some of the limitations of this approach, we present a simplified yet biologically defensible model of the developmental process. The design issues that arise when formulating this model at the molecular, cellular and organismal level are discussed, and for each of these issues we describe how they were resolved in our implementation. We present and analyze some of the morphologies that can be explored using this model, specifically one that has agent-like properties. In addition, we demonstrate that this developmental model can be evolved.