Last modified: 3 Dec 2008 PLEASE IGNORE THIS DOCUMENT The ideas have now been revised and reorganised in this paper presented at IJCAI 2005: http://www.cs.bham.ac.uk/research/projects/cosy/papers/#tr0502 The Altricial-Precocial Spectrum for Robots Aaron Sloman and Jackie Chappell (2005), in Proceedings IJCAI'05, Edinburgh, pp. 1187--1192, Followed later by http://www.cs.bham.ac.uk/research/projects/cosy/papers/#tr0609, Jackie Chappell and Aaron Sloman, Natural and artificial meta-configured altricial information-processing systems, International Journal of Unconventional Computing, 3, 3, 2007, pp. 211--239, and various others. Search for our names Thanks. ======================================================================= Last modified: 31 Jan 2005 (Update above Dec 2013) PRECOCIAL VS ALTRICIAL ROBOTS -- SOME THOUGHTS Aaron Sloman http://www.cs.bham.ac.uk/~axs/ [ Notes and reflections partly based on concerns about the longer term strategy for CoSy and similar projects. http://www.cs.bham.ac.uk/research/projects/cosy/ http://www.cs.bham.ac.uk/research/projects/cosy/PlayMate-start.html These ideas arise partly from discussions with Jackie Chappell. http://jackiechappell.com/ They are also related to discussions of nature vs nurture in relation to theories of 'symbol grounding' discussed in these two (partly overlapping) presentations: http://www.cs.bham.ac.uk/research/cogaff/talks/#meanings http://www.cs.bham.ac.uk/research/cogaff/talks/#grounding ] CONTENTS -- THE PRECOCIAL <---> ALTRICIAL SPECTRUM -- CONJECTURE 1: -- BACKGROUND -- . PRECOCIAL SPECIES -- . ALTRICIAL SPECIES -- . WHY? -- POSSIBLE BENEFITS TO ALTRICIAL SPECIES -- A COMMON ASSUMPTION -- CONJECTURE 2: -- WHAT ARE THE MECHANISMS? -- CONJECTURE 3: -- CONJECTURE 4: the origins of syntax -- CONJECTURE 5: layering -- SUMMARY -- THE PRECOCIAL <---> ALTRICIAL SPECTRUM Suggestion: in order to address the nature/nurture issues raised in our proposal we need to understand how precocial and altricial species differ. We may need to come up with a properly analysed taxonomy of types of robots spread along the spectrum, explaining why precocial robots may be better suited to some tasks and environments and altricial robots to others. This will require us to understand how the architectures, mechanisms, forms or representation, and types of learning differ between the two extremes. As far as I know this has never been done adequately, and the most widely believed theories about the altricial end of the spectrum are false. -- CONJECTURE 1: There is not a sharp distinction, but some sort of spectrum of cases, with many discontinuities, and perhaps a partial ordering rather than a total ordering. -- BACKGROUND -- . PRECOCIAL SPECIES (deer, horses, chickens, and many others) are born or hatched relatively well developed and competent, e.g. able to walk, or feed themselves. Some deer can run with the herd within minutes of birth. So a huge amount of information about structures in the world, how to perceive them, and how to behave in relation to them must be innate in those species. -- . ALTRICIAL SPECIES (lions and other hunting mammals, apes, humans, crows, eagles etc.) are born relatively helpless and incompetent, but by the time they are adults they are generally able to perform tasks of far greater cognitive complexity than precocial species, and in some cases (e.g. humans) in situations with far greater diversity. In some cases the competence could not have been genetically selected for because the environment had never previously existed. -- . WHY? There are heavy costs to altricial species: because the neonates are so helpless adults have to commit resources to guarding, feeding, and in some cases training/teaching them, possibly over extended periods. This can both endanger the adults (make them more vulnerable to predation, deprive them of food given to the offspring, and reduce breeding frequency). (Compare r/K evolutionary strategies, many cheap offspring vs few expensive offspring.) -- POSSIBLE BENEFITS TO ALTRICIAL SPECIES There may be several different biological advantages gained by the 'altricial strategy' as well as costs, e.g. if it helps to overcome some or all of the following problems, each of which can be seen as a resource limitation. - the capabilities required in adults of some species may be too rich to be encoded in the genome (some kind of 'space' limitation). (Genetic information capacity as a scarce resource) - the evolutionary history of some species may not have provided any contexts in which certain currently useful capabilities could have been selected for (NB: evolutionary history is a biological resource, and may have its limitations like any other biological resource.) - slower forms of learning (gradual, adaptive, shaping, processes, as in reinforcement learning) may not be able to cope with the very varied environments and challenges facing adults of some species. (Time to learn as a limited resource) - it may also be that in order to cope with a new problem an individual may have to produce a discontinuous change from previous behaviours, i.e. novel, creative behaviour, as opposed to interpolating or extrapolating in a space that is already well explored. That's because prior learning does not always take the individual 'close' to problems that can occur. (Individual learning opportunities as a limited resource) All this points to the need for a kind of learning, decision-making, and acting capability that supports substantial discontinuous changes from what has gone before (either in the evolution or the species, or in the learning by the individual) without those changes being disastrous. -- A COMMON ASSUMPTION A common assumption among some AI researchers (and others) is that something like human adult intelligence could be a product of neonates born with something like a complete architecture, devoid of all knowledge, but possessing drives and needs and a powerful general-purpose learning mechanism, e.g. reinforcement learning driven by drives, needs, and aversions and results of prior actions, with very little innate knowledge. Then all knowledge would gradually be built up by continual shaping of internal and external responses to various combinations of internal and external stimuli. A possible objection is that this essentially ignores the fact that evolution can, and often does, provide huge amounts of information (in the form of knowledge and skills) that can be used by members of a species acting either mostly in isolation (e.g. pandas? spiders?) or in social structures (e.g. termites) to achieve things that learning would not be able to produce in the lifetime of the individual. So why should that strategy not be used by all species, and by all intelligent robots? Why aren't all intelligent species precocial, i.e. born or hatched highly competent and knowledgeable. We know that in principle this is possible: newly hatched chickens don't have to learn to walk around and peck for food and newly born deer can find their way to their mothers' nipples to feed, and can run with the herd, within minutes of being born. Could not the more intelligent species combine being precocial, i.e. born with a great deal of competence, with being adaptive, i.e. able to continually adapt through reinforcement learning which can gradually vary how the system works, including adding new chained responses? -- CONJECTURE 2: The answer is that evolution 'discovered' and deployed the power of an architecture and developmental strategy towards the more sophisticated (altricial) end of the precocial-altricial spectrum. This strategy includes mechanisms for bootstrapping a wider variety of competences through interactions with the environment during the final processes of physical development of the brain, and beyond. This may explain the creative problem-solving of crows, the ability of some social animals to absorb a culture, and, in humans, the ability to learn and use a rich and highly structured language. -- WHAT ARE THE MECHANISMS? -- CONJECTURE 3: Altricial mechanisms are NOT driven by changing biological needs and desires (e.g. for food, drink, shelter, escape, mating, etc.) concerned with physical survival and well-being, using rewards and punishments to drive change. Instead their key feature is constant experimentation with internal and external actions during which they learn *chunks* of information about what can be done and what can occur. This includes both chunking passively perceived inputs (e.g. using something like Kohonen nets or other self-organising classifiers to chunk sensory inputs) and also discovering that output signals can be chunked into different categories according to how sensed results change. The number and variety of such sensory and motor chunks will depend on (a) the innate physical design of the organism or robot (e.g. having limbs or digits, or jaws, or tongue, or neck, head, eyeballs, ears, that can move independently, with motion changed in various ways, such as accelerating, decelerating, changing direction etc.) (b) the innate collection of internal operations on data-structures that can be performed independently. The system may have some innate control system that performs varied internal and external operations driving these mechanisms, as well as innate mechanisms for learning, storing, labelling various kinds of chunks. Note that different kinds of bodies and different kinds of initial brain mechanisms will produce different possibilities for such learning, which may vary enormously both in the kinds of chunks that can be learnt and also in the number of different chunks that can be learnt. Having two hands with five fingers each with several (more or less independently controllable) joints can produce huge combinatorial possibilities. Different designs for mouths (or beaks) and tongues or vocal mechanisms may also produce different combinatorial spaces to be explored. So different altricial species will differ in many ways related to these differences. In particular some may have much richer combinatorial competence than others. -- CONJECTURE 4: the origins of syntax Some, but not necessarily all, altricial species can not only store, label, categories input and output 'chunks' that can be re-used later, but may also be able to combine them to form larger chunks that are explored, and if found 'interesting' (according to criteria I'll discuss another time), also stored, labelled, categorised, etc. so that they become available as units for future actions. A simple demonstration of how this can enormously reduce search spaces, inspired by discussions with Oliver Selfridge over 20 years ago can be found in the program described here: http://www.cs.bham.ac.uk/research/poplog/teach/finger That example requires the learning agent to store larger units composed of sequences of varying length [A1 A2 A3 ...] where each of the Ai may be arbitrarily complex and repetitions [REPEAT A] were A may be arbitrarily complex and in the example all such repetitions stop only when a 'terminating' event produced by the environment prevents further repetition. Obviously this very simple syntax is just a special case, with obvious omissions including conditionals for instance. Anyhow the conjecture is that some altricial species have an innate compositional competence which is applied both to the perceptual and to the action chunks to produce larger stored structures, using compositional semantics for descriptive or procedural meaning. -- CONJECTURE 5: layering If the methods of syntactic composition are themselves subject to the same process then the result may be production of ontological layers in many parts of the system including perceptual layering, action layering and various kinds of internal layering of control and description. Different kinds of competence may produce different kinds of layering. Some of the above processes are clearly amenable to cultural influences: e.g. what sorts of toys and games are found in the environment during early development will make a huge difference to what can be developed. Rocks, sand, water, climbable trees, smooth terrain, rugged terrain, flat terrain, sloping terrain, etc. will all make a difference too. The above processes could lead to different kinds of understanding and representation of space, time, motion and causality in different species. It may be that in humans after all the above mechanisms evolved in forms that we share with some other mammals (e.g. chimps, bonobos) typical evolutionary processes of duplication followed by differentiation rapidly produced variants that were particularly suited to what we now call language. Example: the ontology of question types being developed here http://www.cs.bham.ac.uk/~axs/misc/question-ontology.html could be the product of the sort of exploratory/combinatorial learning described here. In humans it seems to take between 10 and 20 years to learn? -- SUMMARY The rapid, automatic, non-need-driven collection/creation of a store of labelled, reusable, perceptual and action chunks, along with 'syntactic' mechanisms for combinatorial extensions of those 'basic' chunks, provides a rich and extendable store of rapidly deployable cognitive resources. It may be that if some of this happens while the brain is still growing the result can be 'compiled' into hardware structures that support further development in powerful ways. (This is very vague.) I am not saying precocial species cannot learn and adapt. It's just that the amount of variation of which they are capable is more restricted and the processes are much slower (e.g. adjusting weights in a neural net trained by reinforcement learning, as contrasted with constructing new re-usable components for combinatorially diverse perceptual and action mechanisms). So I suggest that in addition to the experiments where we pre-design working systems with combinations of different sorts of competence in order to understand ways in which such competences might work and cooperate, we should also begin the longer term exploration of architectures and mechanisms for a robot towards the altricial end of the spectrum described above. ===== Some draft notes on requirements for perceiving affordances related to manipulating structures composed of 3-D objects, here http://www.cs.bham.ac.uk/research/cogaff/challenge.pdf An attack on symbol-grounding theory, and a proposal for symbol-attachment theory can be found here: http://www.cs.bham.ac.uk/research/cogaff/talks/#meanings http://www.cs.bham.ac.uk/research/cogaff/talks/#grounding ===== It is hoped that speakers at this two-day Tutorial at IJCAI-05 in Edinburgh, 30-31 July will shed light on these issues: http://www.cs.bham.ac.uk/research/projects/cosy/conferences/edinburgh-05.html Aaron Sloman