Section >> Self : Research
browser, go home Structural coupling and the emergence of autonomous adaptive behaviour
[Home || Research || Autopoiesis || Papers || Alife + || Misc.Info]  
go to bottom of page
Subsection >> select?
  Theory Overview
  Implementation Biosys : Phenomorph
  Related work GRNs : Structural Coupling / Control Systems : Embodiment : Alexander Riegler : Filippo Menczer : Barry McMullin : Behavioural robotics [Brooks : BEAM]
go to top of page


Autonomy, adaptation and autopoiesis
My PhD research began with the idea of instantiating in software of those aspects of Maturana and Varela's theory of Autopoiesis that concern the behavioural characteristics of living systems—in particular, their autonomy, and their ability to adapt over their own lifetimes to variations in their environmental circumstances. The primary feature of Maturana and Varela's work of particular interest to my current research is the process of 'structural coupling' between a system (an organism, agent, etc.) and its environment. Over ontogenic (an individual's lifetime) and, indirectly, phylogenic (evolutionary) timescales, mutual perturbation between system and environment produces a fit between the structure and dynamics of system and environment such that an observer watching the interactions between them sees coherent and meaningful behaviour exhibited by the system in response to environmental stimuli.

On embodiment and structural coupling
For my thesis, I am focusing on the idea of 'embodiment', actively defining it in terms based on the notion of structural coupling. This idea came about as a result of a number of thought processes. I found it odd that there wasn't (back in 1998 at least) any clear definition of what it means for something to be embodied—despite the apparent significance of embodiment in fields such as behavioural robotics, where the fact of a robot being situated in and responding directly to its physical environment plays a key role in the production of meaningful behaviours. It struck me that although there seems to be something primary about physical embodiment, the significance of being embodied is best expressed in terms of how this conditions, shapes and defines the relationship between the thing that is embodied, and the environment that it is embodied in.

For example, being physically embodied in the world creates a very rich and complex relationship between a robot or biological organism and its environment—but it is this relationship that is important, which in the case of physical entities arises from the interactions between physical things under the laws of physics. The bare fact that a robot or organism is made of physical stuff is not significant for embodiment, except in so far as it impacts on this relationship.

A relational perspective on embodiment fits very neatly with structural coupling, which is a fundamentally relational, ontology agnostic concept that describes the significance of non-destructive mutually perturbatory interactions between two systems over a period of time. I decided that it would therefore be interesting to define embodiment as a situation in which structural coupling between a system (a network of two or more components arbitrarily identified as an entity by an observer) and its environment can occur:

A system X is embodied in an environment E if perturbatory channels exist between the two. That is, X is embodied in E if for every time t at which both X and E exist, some subset of E's possible states have the capacity to perturb X's state, and some subset of X's possible states have the capacity to perturb E's state.

See any of my publications from 1999 if you'd like more detail on this. In a nutshell though, the key points about this definition are:

  • It is defined in terms of the relationship between a system and its environment, such that the stuff of which they are made (atoms, bits, higher-level software constructs, whatever) is not directly relevant;
  • It describes a basic criterion for embodiment, but one which is enormously broad in that most things are embodied (for example, every physical thing is embodied in the physical world) unless you start getting quite creative with what you identify as taking the role of 'system' and 'environment' (for example, two software processes running on two non-networked physically separated battery-powered computers);
  • Things get more interesting when you start to consider the significance of changing the values of the variables in the definition, such as the number of perturbatory channels, their bandwidth, and the structural complexity and plasticity of X and E. In this sense, the definition supports the idea of 'degrees of embodiment'; that some things can be embodied to a greater or lesser extent than other things.


go to top of page


Genetic Regulatory Networks
My current experimental platform, called Biosys (it's a biologically inspired system...) uses a simple Genetic Regulatory Network (GRN) model to produce emergent behaviours through structural coupling between system and environment. There's lots of detail about Biosys in my ECAL 2003 paper, here. Basically, it uses a fairly simple model of genes and regulatory proteins with the added features that proteins determine effector output, sensor input is manifest as proteins, and genomes are evolvable. This protein-based input/output means that the GRN (and in particular, its regulatory dynamics) can be used as a real-time control system in a range of different environments. This model differs from other ALife GRN research, in which GRNs tend to be studied as closed systems, or used to 'grow' virtual agents in which GRNs play no part once the agent has been produced (what I call 'disposable genome syndrome'—they certainly aren't disposable in real living organisms).

On embodied GRNs
Biosys uses GRN dynamics as the basis for sensory-motor correlations giving rise to emergent behavioural strategies in relation to the system's environment. In this sense such a system is very much embodied, coupled to an environment via its sensory and effector surfaces, relying on the dynamics of its structure within its environment as a basis for the generation of observable behaviours.

The system's actions in any particular situation are not directly programmed or prescribed. What the system does at any moment in time via its effector surface is determined by its structure at that moment in time, linking sensory and effector surfaces. The structure is affected by both sensory surface (environmental) events and internal dynamics, allowing structural coupling between system and environment to occur.

Behavioural strategies
The thermostat is a very simple example of a simple adaptive behavioural strategy in an artificial system, giving rise, from the perspective of an observer, to 'temperature regulating' behaviour. Chemotactic bacteria, such as E. Coli use a similar homeostatic strategy as a basis for 'food seeking' behaviour, via a 'run-and-tumble' strategy that arises because of the way nutrients in the environment of a bacterium affect its structure, in turn determining the direction that its flagella rotate, this in turn affecting the movement of the bacterium within its environment.

Initial experiments evolving small environmentally-coupled GRNs shows similar reactive behaviour strategies emerging. The following graphs plot the movement of a simulated Khepera robot (Webots v2) over a 0.6 metre square flat surface featuring nothing but a single light-source. The robot is controlled in real-time by an evolved 5-gene GRN with a single regulatory site on each gene. The genome is evolving to respond to protein '0' produced by light detected on the left side of the robot and protein '1' produced by light on the right, in such as way as to maximise exposure to light by producing appropriate amounts of protein '6' (which determines the speed of the robot's left wheel) and protein '7' (determines speed of the right wheel). Again, a more detailed description of the model can be found here.

  Phenomorph Phenomorph, mentioned in my 1999 papers, was an early attempt to investigate the relationship between system-environment coupling and emergent behavioural strategies. The idea was to use Cellular Automata (CA) or Boolean Networks (BN) coupled to a range of different environments (the 'real' physical world, a simulated physical world, the 'real' World Wide Web and a local hyperlinked HTML file repository) via appropriate sensory and effector surfaces to produce emergent chemotactic behaviour. The 'structural coupling to different environments' part wasn't too hard, but trying to get a 'standard' CA or BN to produce dynamics suitable for the emergence of meaningful behaviours was exceedingly frustrating, and so I turned to Genetic Regulatory Networks...
  Related work
go to top of page

The GRN model I use is similar in some respects to the one developed by Torsten Reil.

Peter Eggenberger uses GRNs as part of a development process to produce morphologies, with emphasis on biological realism; as does...

Sanjeev Kumar at UCL, who focuses on phenomena relating to pattern formation.

Josh Bongard uses GRNs as part of a development process to produce artificial agents.]

Wolfgang Banzhaf is doing work on modeling GRNs and their dynamics.

Stuart Kauffman thought of modeling GRNs before anyone else did.

Hidde de Jong, working for the HELIX project, has a biologically grounded GRN regulatory network model called GRN.

Ron Weiss (man of many home pages: here, here and here) uses real cellular circuitry for control systems, though focusing on formal computation.

  Structural coupling / control systems

A number of people have developed systems and models that exploit system-environment coupling to produce control systems, though they don't necessarily talk in terms of 'structural coupling'.

Frank Daellart and Randall Beer have co-authored many interesting papers, often concerned with the coupling between neural network-driven agents and their environment, from a dynamical systems perspective.

Jens Ziegler (a man not afraid to use Java applets for site navigation) has evolved artificial chemistries, similar to biological signal transduction pathways, for robot control.


Tom Ziemke is quite the embodied man-about-town these days :-)

Rolf Pfeifer does lots of embodiment-related work, with a strong 'embodiment == being a physical thing' bias (well, from my perspective it's a bias).

Esther Thelen approaches Cognitive Science from the perspective of embodiment.

Alexander Riegler has an interesting article on embodiment available for download here.


Alexander Riegler

[Radical Constructivism]

Alexander Riegler, at the Free University of Brussels, works on what he calls Constructivist ALife. His perspective is fairly similar to mine - recognising behaviour as a relational construct, and developing systems that exhibit emergent relationships with their environments as a result of their constituent plastic structures that perform sensory-effector correlations. However, the 'guts' of his system are very different, reflecting his background in cognitive science. His interest in Maturana and Varela's work is based on their contribution to the constructivist perspective.

  Filippo Menczer

Filippo Menczer's work in Genetic Algorithms recognises the role of complexity in the environment in the emergence of complex behaviours in the systems that evolve in those environments.

His approach is based on characterising problem spaces in terms of energy ('Latent Energy Environments'- LEE). Evolved organisms must maintain enough energy to continue their existence. Two consequences of this approach are, first, that fitness functions do not need to be imposed from outside of the system in an a priori fashion as is usually the case (this is an 'endogenous' fitness paradigm), and second, otherwise incommensurable environments can be compared via articulation in terms of latent energy. Menczer also eschews normalisation of fitness measures in populations, in favour of a more natural 'local fitness' model, well suited to varied environments.

InfoSpiders is an example application that illustrates the practical benefits of Menczer's approach, as a basis for distributed information seeking on the WWW. A population of agents get the energy they need to survive by finding web pages containing desired keywords. Energy is expended following URLs. This has similarities with the bacterial 'run-and-tumble' strategy, in that the result of this is that positive 'food' gradients in the environment are pursued.

  Barry McMullin
Barry McMullin's work focuses on formalising our understanding of Autopoiesis as a process — currently somewhat 'wooly', and decidedly descriptive in nature. Using the Santa Fe Institute's SWARM simulation system, Barry re-implemented and has extended Varela's initial work.

Varela's initial and McMullin's subsequent simulations are based on the simulation of a qualitative chemistry, describing a range of basic elements, and rules for interactions between them. The goal is to understand the processes by which self-maintaining bounded structures (or just boundaries) arise.

  Behavioural robotics

Generally, embodied approaches to robotics exploit the relationship between robots and their environments to produce meaningful behaviours, rather than trying to stuff robots full of cumbersome and inevitably inadequate representations of their environments. The environment already does a perfectly good job of presenting itself - why re-present it? A grounding in physical reality is prioritised over abstract symbol manipulation. This approach has an awful lot in common with the European School of Philosophy (in contrast to the Western Analytic tradition, exemplified in symbolic AI) - particularly the phenomenology of Martin Heidegger and Maurice Merleau-Ponty. This perspective also motives some of the work being done at COGS in Sussex (see in particular EASy - the Evolutionary and Adaptive Systems Group, more links here.)

Such approaches present human beings as being fundamentally 'in the world', in contrast to the mind-body dualism presented by the father of Modernism, Rene Descartes—which begins with rational thought, and then contemplates its relationship to the material world. This is a direct analog to symbolic AI, which takes formal, abstracted rationality and the mechanisms thereof as its starting point in the development of systems that relate meaningfully to the world and its inhabitants.

Rodney Brooks
[Homepage] [Publications] [Online publications]
Rodney Brooks (who leads MIT's mobile robots group) is generally considered to be the father of modern behavioural robotics - although the infrequently recognised Grey Walter actually pioneered the approach, even if he didn't put the same label on his work. Brooks robots are based on layers of simple functional units - 'move', 'turn' and so on - the most basic functional layer contribute directly to behaviour. Complex behaviours emerge from the combined behaviours of the various functional units within and in response to some environment.

BEAM Robotics (Biology, Electronics, Aesthetics, and Mechanics)
[BEAM Links and Resources] ['Official' Home site]

BEAM robots, pioneered by Mark Tilden, effectively take Brook's approach to an extreme—with a great deal of success.

BEAM robots are constructed solely from simple electronic components, such as resistors, capacitors and transistors. Their behaviours effectively arise as a result of the sensory-effector correlations performed by these simple components. They are totally autonomous, in the sense that they are guided only by their own structure and the dynamics therein, and typically have their own means of generating power, in the form of solar cells.

BEAM robots are extremely cheap to build (allegedly around £20 for the parts for a simple 'SunEater'), placing this technology within the reach of thousands of hobbyists, speeding the evolution of this field...

Tilden has also developed a basic but functional nervous system (a 'nervous net') allowing complex behaviours, but employing just a handful of basic components connected in a ring.



Home || Research || Autopoiesis || Papers || Alife + || Misc.Info
go to top of page

$Id: research.html,v 1.4 2003/06/05 16:09:52 ucactqu Exp $