I want to examine several examples of Artificial Life (AL) systems in which sensing and therefore the requisite sensors are used as a functional part of the system, ie used with intentionality on the part of the system designer and with intentionality on the part of system "creatures". 

It is one of the basic concepts of AL that it involves the simulation in computing systems, robots and other hardware/software configurations, of various functions which operate within, so to speak, "normal" biological life. These processes usually consist in the simulation of evolutionary processes and the use of genetic algorithms to develop independently behaving, or autonomous, entitities which are able to carry out some subset of the possible distinguishing characteristics and processes of living systems. These often include processes of sensing, feeding (or other survival maintenance), mating and reproduction, mobility and the like. The entities being considered are usually some kind of cellular automaton which has some of the characteristics which make up the idea of "agenthood".

In examining sensing in AL systems I want to explore the role of sensory processes and their extension into the areas of comunication between agents and the development of intentionality within an agent. The processes of sensing, communication and intentionality all contribute greatly to the question of the actual assessment of the success of AL projects in their implementation of aims such as the emulation and recapitulation of the evolution of biological organisms from bacteria to humans.

Ultimately, for the purposes of any autonomous "biological" agent it is the processing of information that is the primary motive in sensing the the organism's context and in the organism's engaging in communication. Definitions of my use of these terms are in order:

Information is what is carried in a physically embodied (or simulated version thereof) difference relation generated by aspects of an organism or coherent entity within the context of some system. This concept of information is derived from Bateson [Bateson, 1973] where, in the context or environment of the "organism", it is news of a difference and within the "living" system it is the difference which makes a difference. [for a discussion of information and its embodiement see Jones, 1999]

Sensing amounts to a coherent system's (ie. an "organism's") capacity to absorb difference relations from its context and to carry out such transforms (processing) of those sense-data as to have them available as information about that context.

Communication is the excretion (outputting) of a probe into the context in order to elicit a sensible response from that context. When sensible to another entity, which may or may not respond, a communication between "creatures" can occur (is enabled).

Intentionality may be said to appear when the sensory and/or communicative act is produced in the "direction" of an object in the context for the specific purpose (intention) of eliciting information from the object.

Given that the intention of AL experimentation involves the emulation of biological processes I suggest that the concepts of an organism and a coherent system or autonomous agent have sufficient similarity for them to be reasonably interchangeable within this discussion. In fact, I am arguing that the issue of sensory and communcation systems and the entity's intentionality is a guiding criterion for establishing the success of, and understanding areas of failure in, artificial life projects.

Perhaps more strongly I would argue that the use of sensory and communicative processes by  autonomous agents may be a more fundamental criterion than the generation of such agents or organisms by evolutionary processes, given that it is not yet proven that algorithmic top-down type processes must fail to produce AL organisms.

So having said that it is incumbent upon me to demonstrate what would be a sensory-cum-communicative process at the lowest possible level and to indicate where intentionality might become relevant. I will then look at some examples from a by-no-means-exhaustive review of the literature.

For all practical purposes the single cell is the lowest level unit of coherent systems which demonstrates the basic attributes of a living system. A virus, though more minimal in level, fails to demonstrate at least one generally considered necessary aspect of living systems and that is independent rather than parasitic reproduction, in that it requires the agency of a host's genetic transcription services for its replication. (Of course what this implies about sexual reproduction I'd rather not go into here.) 

In cellular behavior, whether as a single celled creature operating independently within some biochemically active environment or as one of myriad cells in some large scale organism, the process of being able to sample and sense the chemical soup in which the cell operates is a basic necessity. This is the only way it can acquire the basic chemical feedstocks for protein synthesis of any and all of the molecules its system requires for whatever processes it has evolved to exercise. So the activity of opening the cell membrane to transport a molecule from outside to inside though essentially a transport function is in fact a primary sensory act. The shape of the membrane pore may well be such as to only recognise one particular molecule but this is of no consequence since in the high level of intelligent animals the sensory pore (or port) known as the eye is, itself, designed specifically to recognise only a very limited range of but one kind of physical particle, the photon of visible light.

If the cell, as is most likely, should excrete, via a membrane transport process, some kind of chemical (by)product of its internal processes and should some other cell find that that molecule has some kind of significance for it, that is, is recognised by some membrane pore, or is sensed by that other cell, then a communicative act has occurred, minimal as this may seem in this case. When looked at in the context of, say, neuronal cells, this process then becomes highly significant and the communicative function utterly clear. This is exactly synaptic transmission and it is what gives us all the capacity to understand and operate within the world, our chemical and cultural soup. [For a detailed discussion of the sensory and informational aspects of the biochemistry of synaptic transmission and nervous system activity in general see Black, 1991]

So it can be seen here that communication is a sensing process augmented by an output process which within the context produces two-way interaction. Now, at this level it is probably specious to suggest that this kind of communication demonstrates any form of intentionality, but I would suggest that autocatalytic interaction with the chemical environment of the cell and the mutual development of cells in general would have been a necessary part of the process of the evolution of pre-biotic systems, just as this bootstrapping is a necessary part of all biological systems. And of course, as an aside, the production of diversity has an essential role here in the maintenance of the overall system as originally pointed out by Ashby in his development of the idea of ultrastability. [Ashby, 1952]

Intentionality appears when the cell or agent (perhaps by now, necessarily a multicellar organism, though it is by no means proven that single cells do not show intentionality given some of the more interesting aspects of cell behaviour in being able to modify its chemistry to accommodate serious shifts in the content of its environmental soup) actively produces a probe into its environment in order to elicit some difference relation and potential information from some recognised aspect thereof. In other words the sensing now has to be for the "cell's" own autonomous purposes, whether it wants to know who's around the district or if it's good to eat. Autonomous behavior depends on intentionality even if this is highly stripped back.

Craig Reynold's Boids
Boids are a set of autonomously behaving virtual birds - boids ( code objects) in a common contextual space (of computer memory) - which are aware of other similar code objects sharing the same context. Essentially each boid in the flock is a distributed processing element in the system. They are able to behave as a flock of apparently co-ordinated entities on the basis of three rules:

1. 1: avoidance of collisions with neighbours
2. 2: matching velocity (speed and direction) with neighbours
3. 3: maintaining a local centering within the flock by staying close to neighbours.

The necessary information that the boids need for navigating according to these rules is provided by a perceptual model "which makes available to the behaviour model approximately the same information that is available to a real animal as the end result of its perceptual and cognitive processes." [Reynolds, 1987, my emphasis.] This simulation does not, therefore, involve sensing in a way consistent with what we normally think of as sensing. In fact it bypasses any sensing at all and provides what amounts to the output of the sensing system directly to the behavior model in the boid. This output is a filtered selection of the available data in the context memory, providing position, orientation and velocity data with the sensitivity of data about the neighbors decreasing exponentially with distance from each particular boid.

Thus no actual sensing takes place but the equivalent result of sensing is provided directly to the brains of each boid so that they in effect have senses and their behavior is directly dependent on this "virtual" sensing substitute. It would not, in my opinion, be overly charitable, to say that this is perfectly good sensing within the abstract way that I am discussing the term. The boids have available to them the sense-data as information that the system provides at each time-step of its process, and so given the distributed "brains" of each boid in the context space, each cell of the system is sensing the difference relations available in its environment. I don't think, however that this legitimises some claim that communication is taking place, and I don't think that Reynolds is making such a claim at all. Further, since the sensing process is a programmer-imposed aspect of the boids behavioral routine occurring regularly with each time-step of the system's execution any claim of intentionality could be at best only borderline.

Werner and Dyer's experiments in the evolution of Communication
Werner and Dyer used a toroidal (wrap-around) grid of 200 x 200 squares as an environment in which to evolve a male and a female version of the same genotype to explore the possibile evolution of communication between the two types. In this purposefully restricted situation the "animal's" task was to find mates. The females could only emit sounds in response to sighting a male within a close region around them and the males could only hear the signals and move on the basis of their interpretations of "instructions" contained in those signals. This way the environment exerted some pressure to evolve communication rather than having an external operater impose their view of how the communcation might work. As they put it "To put evolutionary pressure on the animals to communicate, we needeed to design animals in an environment such that some animals would have information that other animals needed to know but were not capable of finding out for themselves. The animals with this valuable information would have to communicate it to the other animals." [Werner and Dyer, 1991, p661]

The signal generation and the interpretation of instructions were conducted by recurrent neural nets. When the female "sighted" a male the activation of the input layer flowed through to produce a signal at the output layer which could be interpreted by the male as an instruction to move in a particular direction, (forward, left or right). The male's task was to find the female for mating. The strategies developed - ie. the neural net weightings - are completely determined by the genome of each creature so that any signal produced or any interpretation of a signal is a function of the evolved genome of that creature. Modifications of strategies occurred through mutation and crossover in the sexual reproduction of the parental genomes.

Sensing is built in and communication is seen to evolve, but because the signalling and interpretations are deterministic this is communication without intentionality as Werner and Dyer point out. The situation is strictly behaviouristic, a stimulus (a result of a sensing) producing a fixed response which may or may not be appropriate to the mating task. The development of good strategies was effected by the simple process of removing aging genotypes given a fixed population size in the environment. If a mating could not take place then there would be no survival of that genome.

Terzopoulos, Rabie and Grzeszczuk (1996) Artificial Fishes
In Terzopoulos, Rabie and Grzeszczuk's [1996] artificial aquatic environment their Artificial Fishes "navigate purposefully through their world by analyzing the retinal image streams acquired by their eyes". These virtual eyes are able to focus on an object in the environment by foveation and controlling the direction of gaze, bringing an object from low resolution peripheral vision into the foveal centre for high resolution vision. Those objects which are focussed upon are externally designated "prey" fishes, other objects not so designated are ignored. Here the sensing is done to read the environment for a particular intentional act, namely hunting, but the intentionality is imposed from outside - as a set of preset fixed habits such as hunger, fear and mating - rather than acquired through some sort of evolutionary selection pressure. 

Importantly the information produced by the sensing process has direct operational use or effect in the creature for the purpose of muscle control both in directing the gaze and in effecting motion control through the muscle actuation system. Terzopoulos et al also use something approaching internal or proprioceptive sensing to enable the Artificial Fishes to actually learn to swim in a productive way, say for hunting. This amounts to an iterative updating of parameters taken directly from muscle driver functions optimized against discrete objectives. Though much of the learning of specific muscle control algorithms is done in the Fish's "brain" Terzopoulos et al intervene in the process to promote controller objectives with certain externally desired characteristics thus guiding the optimization of the motion control algorithms.

Terzopoulos et al claim that the fishes navigate "purposefully" through their aquatic environment and purpose can readily be seen to have a level of intentionality therein, but it isn't easy to distinguish the degree of autonomous agency involved in their Artificial Fishes. Much of the algorithms used are built externally from various optimization functions and then imposed into the fish. For example they arbitrarily chose to use an approximation of the primate vision system in their fish's vision guidance system. But, of course, the evolution of autonomous agency isn't exactly their interest in developing this system. 

Tom Ray, Tierra
Tierra is a system of digital organisms or cells (machine-code programs) which live in a purpose built environment provided by the Tierra Operating System. Networked Tierra is a version of Tierra in which the cells "are able to sense conditions on other machines and move between machines". As Ray points out the objective of the project "is not to create a digital model of organic life, but rather to use organic life as a model on which to base our better design of digital evolution." [Ray and Hart, 1998]

In Networked Tierra each of the organisms (machine-code cpu-structures) starts with a pair of differentiated "tissues" in the ancestor cell/dna. The two tissues provide a reproductive system and a sensory system. Over a sequence of reproductions the cells may lose or modify their tissues. The main purpose of the sensory tissue is to probe other Tierran ecologies on other networked machines to enable a mother cell to determine which of the probed Tierran environments is the best one to send its daughter cell. The sensory tissue sends a "tping" code (a Tierran "ping" command) out to 15 other networked machines and, by comparison of the data returned by the Tierran operating system in each machine, makes a decision as to the currently best environment to which to migrate the daughter cell. Examples of the kinds of criteria the mother cell uses for its determination are speed of system per number of cells in system and average reproduction rate.

The cells in this system clearly demonstrate intentionality in their sensory process but equally clearly they are completely unaware of any other cells in the system and communicate only with the environment. Thus there is no communcative process simply a sensory activity. One could not make claims towards the emulation of living processes here and as Ray makes quite clear his intention is not such emulation but simply the exploration of evolution as a means for developing programs.
 

Sommerer and Mignonneau: Pico_Scan
Pico_Scan is an interactive installation artwork built at ATR near Kyoto. The work is based on graphical elements each of which is an artificial organism in itself. These are displayed on video screens and the audience can view the results and interact with the process through the use of a purpose-built scanning device.

Genetic algorithms sense the environment through the scanner and data produced by the scanner is used by these routines as parameters for graphical displays which they generate. These displays are modified on the basis of distance of an object from the camera/scanner, overall colour of the scanned surface, absolute position in the physical environment and the volumetric body capacitance of the person holding the scanner. These data sources provide the graphic production routines with some sense of the audience in the installation space, their engagement with the work in using the scanners and where they are in that space. Thus the system senses what is going on around it and the audience for the work is integrated into the complex graphical productions of its virtual-life processes.

This is the only example I give of an AL system which is capable of sensing its physical context rather than simply the data or memory space in which it operates. To what extent there is internal sensing and communication between cells generating images I am not privvy, but I suspect the cells are quite independent, simply evolving in isolation according to the use of the scanners by outside agents. Of course Sommerer and Mignonneau's intention is to produce a pleasing visual experience in which the audience has a major role in the creative process. This the graphical results of Pico_Scan most certainly provide. [for a report of their other works see Sommerer and Mignonneau, 1997]
 

If the work of the digital designer is in any way to produce developed evolved digital entities which either resemble living systems or interact with actual living people, then sensory processing and some level of communicative activity is essential to their being. Mignonneau and Sommerer's works, as interactive art works, are among very few AL works which express any interest in the human sphere. All their works show a desire for direct interaction with the audience for the work. [Sommerer and Mignonneau, 1997]. Other works which do elicit direct human interaction in the evolutionary process include Karl Sims' Genetic Images in which the audience has direct involvement through aesthetic judgement in the provision of selection pressure in the evolution of the images. [discussed in Ray, 1997]. 

Although the other works I have discussed here may produce interesting visual output, especially Reynolds' Boids, the primary motivation is in the further development of programming technologies and the visual output is largely artifactual. Thus the kinds of criteria I am drawing out are not operating in these works. These criteria, that the AL system should show some level of sensory and possibly communicative function as well as intentionality if there is to be any autonomy in the "organism", aim towards bringing the productions of AL evolutionary processes into contact with human processes and activities so that the systems do not grow in isolation from the greater environment in which we all dwell. I suggest that to produce creatures which exist in some sort of isolation is in the long run to fail the creatures themselves and force them into a kind of confinement which we, under the same conditions, would consider cruel and unusual punishment. In other words no person is an island, to become human requires the presence of other humans in the bootstrap into the living cultural world and in the long run it would be bizarre and anti-life to insist on complete isolation for any created entity. The kinds of entity that we might wish to keep at bay in this sense are in general produced under exactly those conditions of isolation in which we would, and they must surely, become psychotic and sociopathic. Though the thought of this kind of future development might seem far fetched it is probably good prophylaxis to hold these matters in mind in the early stages of our attempts at artificial creation.

If the enhancement and humanisation of communications and the communicative technologies is our goal then to do it under "de-humanised" conditions surely will breed failure.

But the real point about these kinds of criteria is that autonomy implies self-regulation and self-regulation is impossible without some sort of knowledge, not simply about the "organism's" internal processes, but also about the consequences of these processes on the environment in which the organism operates. In other words the organism has to sense the state of the environment in order to gain information from that environment which may directly effect its continued existence. This is especially so as the organism, and others in the same context, will be producing outputs which effect the environment and those other entities in that environment, thus effecting fitness criteria and the fit of the organism within that context. To know of what you are doing to the environment is a basic sensory process and to know of others in the environment is a basic communicative process. As an autonomous agent, if your activity in that environment is then effecting the state of others also in that space then it is a communicative act to sense the output of the other entities so as to be able to self-regulate, whether that process is essentially competitive or collaborative. This self-regulatory loop is a further basic characteristic of a living system wherein the dynamics of the system have to be brought into consideration by the evolving entity for the simple purpose of self-maintenance and the continued functionality of the environment. Selection pressure must depend on the dynamics of organism behavioiur and Ray and Hart [1997] provide an excellent example of the failures that can result if this condition is not recognised in their comments about "mob behaviour" when one of the available niches in the environment is vastly superior to the rest and all organisms attempt to, in the case of networked Tierra, "migrate" to that most attractive context.

Sensing what is in the environment and communication with other like creatures in that environment are both essential aspects of even the most primitive objects of bio-logic. In any simulation of bio-logic and evolution as a process both sensory and communicative processes are a necessary inclusion. Further, intentionality is a necessary aspect of the autonomous behaviour of any entities in biological and in simulated environments. Though much of the work in AL is more about programming technologies than about the exploration of living systems where there is a crossover these three factors should be kept in sight.

Ashby, W.R. (1952) Design for a Brain. London: Chapman & Hall.

Bateson, G. (1973) Steps to an Ecology of Mind. Paladin Books.

Black, I.B. (1991) Information in the Brain. A Molecular Perspective. MIT Press, Cambridge, Mass.

Jones, S. (1999) "On Qualia and Information" paper read at the Towards a Science of Consciousness conference held in Tokyo, 1999. submitted to Journal of Consciousness Studies (current status unknown). available here

Ray, T.S. (1997) "Evolution as Artist" in: Art@Science, Sommerer C., Mignonneau L. (eds), pp81-91. Springer Vienna/New York. 

Ray, T.S. and 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.], pp295-304. The MIT Press, Cambridge. <http://www.hip.atr.co.jp/~ray/>

Reynolds, C.W. (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model" Computer Graphics 21, (4) Siggraph '87 Conference Proceedings and in Siggraph 1996 Course Notes #36 Artificial Life for Graphics, Animation, Multimedia and Virtual Reality.

Sommerer, C. and Mignonneau, L. (1997) "Art as a Living System" in: Art@Science, Sommerer C., Mignonneau L. (eds), pp81-91. Springer Vienna/New York. see also http://www.mic.atr.co.jp/~christa/

Terzopoulos, D., Rabie, T. and Grzeszczuk, R. (1996) "Perception and Learning in Artificial Animals" Artficial Life V. and in Siggraph 1996 Course Notes #36 Artificial Life for Graphics, Animation, Multimedia and Virtual Reality. 

Werner and Dyer, (1991) "Evolution of Communication in Artificial Orgasnisms" in: Langton, Taylor, Farmer and Rasmussen, eds. Artificial Life II. Vol X in Santa Fe Institute Studies in the Sciences of Complexity, Addison Wesley.