One significant reason for this may be, as TC Schneirla pointed out many years ago, that complex biological systems are not merely quantitatively different from simple ones, but have a radical difference in organisation. And if, as Gould has recently argued, that diversity in evolution is not progress per se, and if the majority of biological systems merely express variations of a core of simple adaptive mechanisms which evolution has conserved, it now invites the question of what is complexity and autonomy in complex biological agents and what, if any, are the radically different features of complex systems which are unique to them ?
One traditional answer to such questions, at least as far as human agents are concerned, has come from the second major stance - a representational, cognitive approach operating at the symbol level and which presumes a linguistic agent possessed, as Piaget once put it, of "semiotic instruments and the like". A persistent problem with this second programme is the very adequacy of such characterisation of the end state of human cognitive growth as terminating in a rule governed symbol based executive (Johnson-Laird, 1983). And whatever its merits in this sphere, it is true to say that no effective ontological position has as yet derived from such a stance that has led to what Piaget himself with all his huge investment in genetic epistemology would have desired- a tracing back to the embryology of the mind (Smith, 1993).
A third way. There is now a growing consensus that a new approach needs to be forged which maps the micro-level functions of the brain to the macro level but without presuming the very representational, symbolic/notational devices of our culture which need to be explained. From an ontological perspective, we need to understand where cognitive systems come from (Van Geert, 1994). In this context, there is a growing realisation that many higher levels of functioning are emergent properties deriving from dynamic interaction between subsystems in ways which need to be properly examined ( Hendriks Jansen ( 1996); Ballard et al, 1997; McGonigle, 1987; McGonigle and Chalmers, 1990, 1996, 1997) and in paradigms which enable the investigator to study key features of cognitive adaptation both in isolation and in combination (McGonigle and Chalmers, 1993, 1996). As many of these have become fossilised, as Vygotsky one put it, - or, in AI terms, 'compiled' in human adults - it is also necessary to study cognitive development from an early stage. Critically we need to be able to specify the design primitives of agents, the regulators of cognitive growth, based on the development of paradigms which for the first time will assess the competences of agents to self regulate cognitively over protracted time periods in response to increasing task demands - a fundamental aim which Piaget, for example was never able to realise even in his functionalist phase (Chalmers and McGonigle, 1997).
At the Laboratory for Cognitive Neuroscience and Intelligent Systems at Edinburgh, in a programme based on a complex systems stance deriving from the comparative and developmental psychology of primates, recently extended into cognitive modelling and robotics, we have sought to play our part in this growing international agenda .
In this talk I shall describe a major paradigm shift in the study of complex intelligent systems based on the combinatorial problem and which has enabled us to re-characterise some of the higher forms of cognitive organisation as the result of dynamic regulations by the agent when attempting to achieve the most behaviour for the least resource.
As a core part of the dynamic, it is crucial to specify a task space which can put the agent under progressive pressure to adapt with ever more efficient data reducing, data management strategies. Ever since the outset of experimental work, it has been a problem to establish a metric of task difficulty. This has made cross species and developmental comparison difficult; however, it has also had the effect of making the mapping between 'the world' as constituting adaptive pressure and the task as its exemplification very difficult. In choosing serial ordering and executive control, however, as the competence(s) we assess, this problem has been markedly reduced. First, the ordering of behaviours at the level of language and action is of high ecological validity (Chomsky, 1980; Lashley, 1960 ); second, the serial task affords a ready metric of difficulty in that the greater the number of constituents to order and arrange, the more combinatorially explosive the possible space of sequences which derive from the core constituents. Thus a three constituent sequence can be ordered 6 ways; a five constituent sequence can be ordered in 120 different ways; and a seven item one is already at 5040. To keep the search space within manageable proportions, the agent either has to have relatively hard wired sequences as may be seen in the analysis of fixed action pattern (as in the case of a finite space automaton) or have effective non random selection heuristics which prune out the majority of the sequences in the total search space. How such pruning may be undertaken by complex agents, however, is only beginning to be understood . I will summarise some of our recent studies in this area, and trace the emergence of self-organised, optimised search procedures from specific design primitives unique to complex biological agents.
A recent derivation of this work has been into robotics where one of our current problems is the design and implementation of functional architectures for task level control in the Nomad, our most recent machine .
Dr. Brendan McGonigle
Laboratory for Cognitive neuroscience and Intelligent systems
Level 8 Appleton Tower.
University of Edinburgh