AI AND ALIFE NOTES IN RESPONSE TO A JOURNALIST'S REQUEST FOR AN OVERVIEW Aaron Sloman School of Computer Science The University of Birmingham http://www.cs.bham.ac.uk/~axs DRAFT: Comments Welcome Last updated: 10 Aug 1998 Added more about needs and niche space and further cross-references 20 Dec 1997 (Added more on trajectories in niche space and refs at end to McCarthy & Minsky) This is one of many half(or less)-baked papers and discussion notes in http://www.cs.bham.ac.uk/~axs/misc/ CONTENTS -- WHAT IS ALIFE? -- -- AI and Alife abstract away from details -- -- AI and Alife subsume each other -- -- . Alife subsumes AI -- -- . AI subsumes Alife -- -- Silly disputes -- TRAJECTORIES IN DESIGN SPACE AND NICHE SPACE -- -- Design space -- -- Niche space: explicit and implicit niches -- -- Changing niches -- -- . Niches changed by the individual -- -- . Niches defined by the needs of others -- -- . Niches changed extraneously -- -- . Self-induced niche changes -- -- Niches for artificial and social systems -- -- Individual trajectories in design space -- -- Population trajectories in design space -- -- "Repair" trajectories -- NATURAL, ARTIFICIAL OR BOTH -- -- Modularity aids intelligibility -- IS IT REAL LIFE, REAL INTELLIGENCE, REAL CONSCIOUSNESS? -- See also: -- WHAT IS ALIFE? Alife and AI each can be seen as a superset of the other. I shall explain why. AI explores mechanisms and architectures for systems able to perform many sorts of tasks, including tasks previously performed only by people and animals. These include perception, learning, planning, problem solving, language understanding and generation, game playing, artistic creativity, and many more. AI investigates the general principles required to explain and model both how people and animals work and also possible designs for new types of systems. So the field includes the study of principles relevant to organisms which have not yet evolved but might evolve in principle and machines (both hardware and software) which might be developed in the future. This includes the study of natural and artificial mechanisms that modify themselves, through learning, adaptation and development of various sorts. Alife extends these studies by exploring ways in which POPULATIONS of systems with reproductive capabilities can modify themselves. This includes investigating the dynamics of coexisting, competing and collaborating populations, and the processes by which successive generations develop. -- -- AI and Alife abstract away from details In both AI and Alife, the study of general principles involves abstracting away from various details, for instance: o Both ignore the differences between naturally occurring and artificially developed systems (e.g. even something without a designer can be thought of as a design which could be thought up by different possible designers.) o Both ignore the differences between different sorts of underlying mechanisms (e.g. organic matter vs wires and transistors) o Both ignore the differences between materially embodied systems and those that exist only as "virtual" or "abstract" machines in software systems. (Not all researchers ignore this difference: there are some whose focus of interest is only in embodied mechanisms, but that is simply one type of focus.) o Both ignore the differences between systems that have been directly designed and implemented and those that have somehow produced their own capabilities by evolving or adapting or learning: if the end result is the same we can investigate its properties without bothering about differences in origins. (Note that in this context "virtual" does not contrast with "real" -- virtual machines in computers, e.g. word processors, are real and perform really useful tasks, including causing physical events to occur such as displaying and printing formatted pages of text and images.) -- -- AI and Alife subsume each other -- -- . Alife subsumes AI Alife subsumes AI insofar as the processes studied in Alife include the evolution of various kinds of more or less intelligent systems. So far, despite much hyperbole in the Alife camp, only extremely simple systems have been evolved, generally very much simpler than the more sophisticated systems designed by people working in AI. However it is to be expected that Alife researchers will go on and study the evolution of far more complex systems. This will require enormous amounts of computer power if, for instance, thousands of millions of years of evolution of multiple coexisting species need to be simulated. -- -- . AI subsumes Alife AI subsumes Alife insofar as AI researchers are beginning to explore ways of including evolutionary mechanisms within a single problem solving system (e.g. a package that uses genetic algorithms to produce a design for a circuit). It seems that something like evolutionary mechanisms are used by our immune system, which is a kind of problem solving system. So perhaps other aspects of the development of intelligence in individuals can use evolutionary mechanisms. -- -- Silly disputes This mutual subsumption means that AI and Alife are essentially studying the same thing, but people in the two research communities have different viewpoints, different centres of attention. Many of them are unaware of how their work fits into this broad picture, so they waste time on silly disputes and pronouncements. E.g. it is not uncommon for Alife researchers to pronounce AI to be doomed to failure. This silliness can in part be attributed to blinkered vision, and in part to the pressures of competition for increasingly scarce research funds. Even though each of AI and Alife subsumes the other, the differences in emphasis, and in level of description are real enough. For instance, they tend to explore different sorts of trajectories in design space and niche space, which I'll now explain. Within each field there are also silly disputes, e.g. within AI there are silly disputes concerning: o the use of logical vs non-logical forms of representation, o the use of connectionist vs symbolic mechanisms o top-down vs bottom-up or middle out approaches, o the use of simplied "toy" worlds vs more complex "realistic" problems o the development of explicitly designed systems vs development of self-modifying systems It is sometimes not clearly understood that alternative, complementary approaches are needed if we are to gain a deep understanding of complementary phenomena. As the field of Alife develops it will probably also generate a rich collection of silly internal disputes, including for instance disputes regarding whether evolutionary mechanisms can produce complex systems with modular, intelligible designs, or only unstructured messy solutions to problems. (For more on silly disputes see Minsky's article listed at the end.) I turn now to some of the deep features common to both AI and Alife. -- TRAJECTORIES IN DESIGN SPACE AND NICHE SPACE Many people study systems that learn, adapt, develop, or evolve across generations. I see all of these as part of a larger study of possible "trajectories" in design space and niche space. I'll explain briefly what these spaces are. -- -- Design space "Design space" refers to the class of possible mechanisms and architectures that can in principle exist, including: o *naturally* occurring biological systems, such as plants, many kinds of animals, ecosystems o *artificially* created, explicitly designed systems such as computers, robots, word processors, the internet, office management software, social organisations, the united nations, etc. Design space includes all possible future systems that have not yet evolved and not yet been thought of by anybody, just as the principles of geometry apply to all possible shapes including those that have never been instantiated in physical objects. (That's the main reason why the general notion of a design does not require designs to have designers, just as mathematical facts and concepts do not require mathematicians: many such facts and concepts have not yet been discovered by any mathematicians, and perhaps some of them never will be. Likewise some classes of designs.) In studying *general* principles, we can ignore irrelevant differences between systems that were explicitly designed by a conscious designer, systems produced by evolution, and systems that occur by chance. Two identical systems produced in different ways have the same properties as far as their inclusion in design space is concerned, even though people may, for instance, have sentimental attachments to, or copyright claims on, one and not the other. -- -- Niche space: explicit and implicit niches "Niche space" is a more subtle and complex notion derived both from the biologist's conception of the niche into which a plant or animal fits and from the engineer's conception of a set of requirements specifying what a design is for. Two sorts of insects in the same place can have very different niches (in the sense used by biologists). So a nich is not a geographical location but something more abstract: namely what an engineer might call a collection of requirements and constraints. E.g. for one insect the niche might involve a requirement be able to locate sources of nectar in certain plants, while for another it involve a requirement to be able to catch and eat smaller insects. One may require the ability to fly, and the other the ability to crawl or climb. The niches considered by engineers are explicitly described and can be used to guide and evaluate designs. By contrast, the niches that occur naturally are implicit in the evolutionary pressures that operate on evolving species, and generally depend on a combination of: o the physical properties of the environment o the properties and behaviour of other species that exist in that environment o each organism's own needs. -- -- Changing niches Previously we noted that designs embodied in animals or machines can change over time, through evolution, or individual development and learning. Similarly the niches instantiated in a geographical location can change over time, in different ways. -- -- . Niches changed by the individual For an individual organism the niche can change significantly during its lifetime. For instance if initially it is fed and cared for by adults and later has to fend for itself, its niche, that is the requirements that its design must meet, will change. Likewise if an organism has different needs at different stages in its development, e.g. requiring different kinds of food or shelter at different times, or if it passes through one or more phases when it has a drive to mate, or to achieve supremacy in a group, then its niche changes accordingly. So some aspects of the niche depend primarily on external opportunities, dangers, and constraints, while others depend on the organism's needs, which may develop as the organism's design changes, through innately programmed development or learning. (How to define the notion of "an organism's need" is a tricky philosophical question. Some of the needs may be more realistically attributed to the gene pool propagated through the organism than to the organism itself. Some needs make sense only in the context of prior commitment to design solutions for other needs: e.g. you need a liver because you already have other design features, not because a liver is an absolute requirement for any successful organism with your capabilities. Likewise you may need other things in part because you have a liver. Needs are justified by circular networks of mutual support.) -- -- . Niches defined by the needs of others In some cases, if the organism is used by something else we can consider the needs of the USER as also forming part of a niche against which the organism (and more abstractly the design which it shares with other similar organisms) can be evaluated. For example the needs of a farming community or their customers form part of the niche of cattle and crops, and have played a role in their evolution, through selective breeding. Most artificially evolved machines and software systems at present do not have needs of their own: their niche is defined entirely by requirements of users. However, in future we can expect increasingly autonomous agents to develop. (Some social systems seem to have outlasted the needs they were originally set up to serve, and have developed to serve their own needs, not the needs of any of their members.) -- -- . Niches changed extraneously Changes in climate, number and type of predators and parasites, amount and type of available food, availability of locations for nests or burrows, etc. can all change the requirements on an organism for both individual survival and successful reproduction. Thus some changes in niches of organisms arise simply from contingent changes in the environment which may have nothing specific to do with those organisms: e.g. they are not produced by users of the organisms. However, some of the changes in the niche of a class of organisms C may be produced by reactions of other organisms which prey on or are preyed on by, or are otherwise affected by members of C. I.e. the design of C partially determines the niche of another class of organisms C', which may change to fit that niche better, thereby changing the niche of C, causing C to change. Such causal changes linking designs and niches may include both positive and negative feedback loops. -- -- . Self-induced niche changes Some intelligent organisms (and some unintelligent ones) may change their environments in ways that affect their niches. E.g. walking through a jungle leaving footprints may make it easier for you to find your way home if you develop the ability to recognize your footprints. Building shelters for use in very cold weather or developing the ability to make clothing will alter the requirements for skin covering grown by an animal. Not all self-induced niche changes are beneficial: over-grazing may seriously reduce a food source for instance. More complex and indirect irreversible effects can also occur, e.g erosion of soil from over-grazing. Human self-induced niche changes are the most obvious and spectacular but there are many examples of different kinds, including an enormous variety of nest-building capabilities in insects, birds, mammals and other animals. -- -- Niches for artificial and social systems All the remarks about interactions between organisms and their niches are equally applicable to artificially produced hardware and software machines, social and political systems, ecosystems, etc. In other words, a larger structure with a design and a niche may be made up of smaller systems with their own designs and niches. Thus all the above developments in niches and designs may occur simultaneously at different levels within a single complex system. -- -- Individual trajectories in design space Because a niche can change during an individual's lifetime many organisms and artificial systems have designs which allow them to develop new capabilities, thereby changing their "location" in design space. Thus there are both naturally occurring and and artificially produced trajectories in design space. In biological organisms we have barely begun to understand the variety of forms of self modification, through many forms of growth, learning, adaptation, and cultural influences. Examples are the development of a newly hatched chick into a fully grown hen, a human infant into an adult, a novice musician to a concert performer, a bright teenager into an expert mathematician, and the learning of languages or sporting skills. (Should we include the development of an egg into a chick? There is no sharp boundary between the process of production of a new individual from the genetic material provided by parents and the process of development of the individual with changing needs. There are also cases where the niche doesn't change, only the degree of fit: as in the continuing growth of a tree over many years.) Thus there is a class of trajectories that an individual can follow through design space, with corresponding trajectories in niche space. Call these i-trajectories. Increasingly, nowadays, there are also artefacts that adapt or learn, e.g. adaptive plant control systems, information networks that adapt their routing decisions as loads change, data-mining systems that develop new forms of classification, interfaces that adapt to users, and self tuning software for operating systems. Techniques for doing this in AI include the use of inductive learning, automatic debugging, neural nets, and various forms of evolutionary computation. These are all important long term research topics in AI. -- -- Population trajectories in design space But not all trajectories in design space are possible for an individual. An egg that is capable of developing into a chicken or an egg that is capable of developing into a dinosaur may be incapable of developing into an eagle or a giraffe, no matter how much the environment is altered to pressurise it in that direction. If instead of focusing on individuals we consider POPULATIONS of individuals with reproductive capabilities then we find a much broader class of possible trajectories. Biological evolution provides many examples of trajectories followed by a population across several generations which cannot be traversed by individuals. Populations rush in where individuals cannot tread. We can use "e-trajectories" to refer to trajectories that require evolution of a population. (An e-trajectory will generally be a little like a "smeared" trail in design space compared with an i-trajectory.) The study of these population trajectories is what Alife is primarily about. It attempts to explore all possible mechanisms by which populations can change themselves in response to environmental pressures. This requires a notion of what makes one individual better than another. In Darwinian evolution this is simply reproductive fitness, which can be achieved in different ways by different sorts of organisms. This can be generalised to include all sorts of fitness measures, for instance fitness in solving a particular engineering goal, or winning competitions at agricultural shows or dog shows, using the simple expedient of letting a researcher or farmer or dog breeder decide which individuals in each generation are allowed to produce offspring. In other words, reproductive fitness may be externally manipulated to guide the evolutionary process. This is why Alife is relevant to engineering as well as to the study of naturally occurring and theoretically interesting examples of evolution. -- -- "Repair" trajectories Some kinds of changes in an individual's design are not possible for an individual nor for an evolving population, but are possible if some external agent intervenes, e.g. by "repairing" a faulty machine or software system, or organism (perhaps using surgery, or chemical treatment). For want of a better label let's call the class of trajectories which are not i-trajectories or e-trajectories but can be produced by external intervention "r-trajectories". -- NATURAL, ARTIFICIAL OR BOTH In both AI and Alife researchers can either restrict themselves to studying and modelling the naturally occurring mechanisms and processes, or, as suggested by the use of the word "artificial", they they can restrict themselves to the study of artificial systems including complex electronic circuits, software systems, more or less autonomous robots, etc. Better still, they can be really broad minded and attempt a general study, looking for principles that cover both natural and artificial systems, just as general principles of aerodynamics are equally relevant to the design of flying insects, birds and aeroplanes. Sometimes the study of artificial systems may give us clues as to features of the naturally occurring system that we had not previously noticed. E.g. my colleague Riccardo Poli works on forms of evolutionary computing that do not use simple linear structures or tree structures to represent a genetic specification but rather networks of genetic material. These can sometimes be far more efficient at solving particular problems than the simpler bit strings used in genetic algorithms. Perhaps one day we'll find that natural evolution also uses "higher level" structures as genetic building blocks, for the same reasons as engineers do, i.e. to constrain and speed up the search for good solutions. This is a particular example of the general investigation of classes of evolutionary mechanisms to find out which are most efficient at solving problems. One of the problems in studying very complex systems which have evolved over thousands of millions of years is whether human minds are capable of understanding that degree of complexity. Some have argued that we cannot hope to. I find this an unconvincing form of defeatism, possibly based in part on a desire to divert resources from those who want to investigate explicit design techniques. How might it be possible for human minds to understand very complex systems produced by evolutionary processes without a designer, including for instance human brains? An interesting conjecture is that in order to overcome problems of searching through combinations too numerous even to be explored in evolutionary time, the evolutionary system has been forced to use design principles which are to some extent modular and structured. If so, the results could turn out to be more intelligible than a huge mish-mash of tiny components which just happen to work together to produce an effect. Thus it is hoped that even if we cannot understand all the fine details of the structure of any particular brain, or any particular evolutionary trajectory, there may be large scale developments, at a higher level of abstraction, that we can understand, and explain, and perhaps one day model in artificial systems. -- -- Modularity aids intelligibility An example might be the development of different architectural layers in human-like intelligent systems, corresponding to recent ideas emerging from work in AI and brain science, namely the notion that human like intelligence depends on at least three different coexisting sub-systems namely o An evolutionarily old reactive system, to be found in many animals (including some forms in insects) o A more recently evolved deliberative system capable of creating internal plans and other complex objects and comparing them and selecting them in advance of executing them. o A still more recent and rare ability to turn deliberative mechanisms inwards and base them on information about internal states and processes, including intermediate perceptual processes (corresponding to the notion of qualia). We can use AI techniques to explore the design principles for such multi-layered systems (as we are doing at Birmingham), and we can explore the consequences of those layers, e.g. showing how the three layers generate three different sorts of emotional states, and also how illusory philosophical problems of consciousness will arise for any system that has the third layer. And we can use Alife techniques to investigate how the layers might evolve (since it is unlikely that fossil records will ever give sufficient information about the evolution of brain structures and information processing functionality). Perhaps by considering what is evolvable we can refine our attempts to understand what has evolved, and by considering what has evolved we can refine our attempts to understand the mechanisms and principles of evolution. I find this synthesis more exciting and challenging than either the defeatist view that it is all too difficult or the exclusive view that everything has to be done in a particular way, e.g. either using explicit design or letting systems design themselves. -- IS IT REAL LIFE, REAL INTELLIGENCE, REAL CONSCIOUSNESS? Some people would object to the attempt, in both AI and Alife, to ignore the differences between the natural and the artificial, or between physically embodied systems and systems simulated entirely in software. They would claim that the attempt to find common general principles linking the natural and the artificial is misguided, because the artificially produced or evolved will not be the REAL thing, or perhaps the software-only versions will not be the REAL thing. It won't be REAL life, REAL intelligence, REAL perception, REAL planning, REAL consciousness. Likewise, some claim that a system inhabiting only a virtual machine environment implemented in software cannot be an example of REAL life, REAL intelligence, REAL perception, REAL consciousness, etc. The problem with this sort of objection is that it is based on dichotomous thinking. The assumption is that we have concepts like "alive", "conscious" "intelligent" which divide things up into two classes, those which the concept applies to and those which it doesn't apply to. So the assumption is that everything either is alive or it isn't. This is obviously silly with concepts like "house". A house is something that has a collection of features that make it a useful enclosure for its occupants. But there's no well defined subset of those features that form a minimal requirement for something to be a house, so that everything that has those features is a house, and everything else isn't. Rather, "house" is a cluster concept. It corresponds to a cluster of features which in various combinations can make something a house, but with no well defined boundary between the cases that are houses and those that are not, even if there are clear examples of both. Maybe under some conditions you'd regard a rectangular sheet of metal supported by four poles as a house, maybe not. Maybe under some conditions you would call Buckingham palace a house, maybe not. But arguing about over whether something is a REAL house if it doesn't have any walls, or any doors, or if it is as big and complex as a palace, is just silly. The important thing is not to draw boundaries, but to understand that there is a large variety of cases with different combinations of features. We can study the implications of the presence or absence of various features, without worrying whether they make something a REAL house or not. We could, if we wish, give them different names, coined for the purpose of making new distinctions that we have found useful, e.g. "wallfree-house", "palatial-house", etc. The same goes for concepts like "alive", "conscious", etc. These are also cluster concepts, which refer in a partially indeterminate way to collections of features which can be present or absent in different combinations. Some subsets (e.g. the features found in a chicken, or a giraffe) definitely make something alive and other subsets (e.g. the features of a rock) definitely don't. But there are many combinations which we have never previously encountered, and therefore our language has not needed to take decisions about whether they do or do not suffice for being alive. Some of those combinations are found in artificial systems. In particular, over many years AI researchers have been examining ways of implementing artificial systems with combinations of various kinds of abilities, including visual perception, auditory perception, tactile perception, motor control, learning, planning, remembering, discovering new concepts, solving mathematical problems, painting pictures, composing poems and stories, communicating with other natural or artificial systems, acquiring new goals or interests, emotional capabilities, and many more. Arguing over whether such systems, whether they are tangible robot like entities, or software agents in virtual reality environments, REALLY are alive or not, REALLY have mental states or not, is a complete waste of time, for there can be no answer. But we can explore the implications of having different combinations of features and, for some combinations that recur often and are of interest, we can coin new unambiguous names: alive1, alive2, alive3, conscious1, conscious2, conscious3, etc., just as, when we discovered that a chemical element such as carbon could have two isotopes we did not need to waste time arguing over which is REALLY carbon. Instead we call one carbon12 and the other carbon14 (or whatever), and then study their similarities and differences. So instead of arguing over whether the entities studied in Alife, or in AI, are alive or conscious or intelligent, or worrying about where to draw the boundaries between those which REALLY are and those which are not, we can simply note that different more refined versions of our old indefinite concepts can be defined, with different boundaries. Then we can explore the implications of each case, e.g. which regions of niche space it can fit, what the implications of its design are. And we can go on to explore more global processes in which such systems interact with one another and either individuallym or in groups, or across many generations, follow intricate trajectories in design space and niche space. This replaces fruitless philosophical (or theological) debates with productive investigation. Let's get on with the job. There's lots to do. ======================================================================= -- See also: ---------------------------------------------------------- John McCarthy What is Artificial Intelligence? http://www-formal.stanford.edu/jmc/whatisai/whatisai.html McCarthy shares the view of AI as science, not just engineering, but has a more sharply focused range of interests, with an emphasis on the use of logic. However, that can be seen as a component of the research programme sketched above, not something in opposition to it. See also his main web page: http://www-formal.stanford.edu/jmc/ Marvin Minsky "Logical vs. Analogical or Symbolic vs. Connectionist or Neat vs. Scruffy", in Artificial Intelligence at MIT., Expanding Frontiers, Patrick H. Winston (Ed.), Vol 1, MIT Press, 1990. Reprinted in AI Magazine, 1991 ftp://ftp.ai.mit.edu/pub/minsky/SymbolicVs.Connectionist.txt This is one of many papers by Minsky emphasising the need to combine different approaches to the study of mind. See his web page for further pointers: http://www.ai.mit.edu/~minsky/ I have a loosely relevant, half-baked, still growing paper on the evolution of consciousness located in the Cognition and Affect project directory ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect The paper is in postscript and compresssed postscript format in two files: Sloman.consciousness.evolution.ps Sloman.consciousness.evolution.ps.gz Related mostly DRAFT discussion papers attempt to explain the connections between the philosopher's concept of "supervenience" and the engineer's concept of "implementation", and other issues mentioned above: ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/Sloman_iberamia.ps The ``Semantics'' of Evolution: Trajectories and Trade-offs in Design Space and Niche Space. ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/Sloman.supervenience.and.implementation.ps Supervenience and Implementation: Virtual and Physical Machines ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/Sloman.design.and.niche.spaces.ps Design Spaces, Niche Spaces and the ``Hard'' Problem http://www.cs.bham.ac.uk/~axs/misc/aiforschools.html An introduction to AI for school leavers and careers advisers. [end]