Cognitive Systems: Towards an Integration of Symbolic and Sensor-Motor Intelligence?
by Cristiano Castelfranchi
The European Commission has identified Cognitive Systems as one of the priorities for the new generation of research projects to be developed from 2003 to 2008 (http://www.cordis.lu/ist/workprogramme/fp6_workprogramme.htm). The stated objective is to construct physically instantiated or embodied systems that can perceive, understand (the semantics of information conveyed through their perceptual input) and interact with their environment, and evolve in order to achieve human-like performance in activities requiring context-(situation and task) specific knowledge.
ERCIM News has chosen to devote a special issue to this exciting research challenge in order to monitor what is under development in Europe (but not only in Europe), and what is the current status of research and development in this domain.
Over the last twenty years, in the field of Cognitive Science, there has been a real scission between the symbolic paradigm, on the one side, and the heterogeneous army of its enemies: neural nets and connectionism, embodied and situated intelligence, symbol grounding, dynamic and evolutionary approaches, distributed cognition, on the other. In the meantime, important changes have taken place. Taking Artificial Intelligence as an example, we can cite the reconciliation between reactivity and planning with the BDI approach (ie, modelling the mind and its processing in terms of Beliefs, Desires, Intentions), or between a reasoning mind and a open uncertain world; the idea of a decentralised and emerging intelligence among distributed interacting systems; the development of new formalisms for autonomous action and cognitive robotics; or the introduction of so-called 'affective computing'. At the same time, more radical and simplistic attacks, such as the elimination of planning, or the elimination of intelligence as manipulation of internal symbolic representations, or the rejection of any modularity of mind, have been relaxed. Thus, a less ideological confrontation, and perhaps even a reconciliation and integration is now possible. It is clear that Cognitive Systems, able to deal with a physical, embodied and situated environment, but also endowed with mental representations, able to use language, and to exploit both experience (learning) and inference (reasoning), reactions and planning, will be a privileged area for such an attempt.
Indeed, Cognitive Systems is a really important research area from both scientific and technological perspectives, dealing with physical environments and mental representations, and exploiting both experience (learning) and inference (reasoning) in order to be able to autonomously explore, learn, react, decide, solve problems, coordinate with each other, interact with humans, etc. There are numerous and important applications under way, from space exploration to emergency handling, from industry to domestic assistance, to worrying military applications. However, all existing projects and working systems are partial and incomplete when compared with the ambitious model targeted in the Call of the Sixth Framework Programme. The aim of this special number of ERCIM News was thus to provide a broad and sufficiently representative view of current advanced attempts to build embodied Cognitive Systems.
Systems can be incomplete in different ways and there are two partially independent paths towards completeness aiming alternatively at:
a) obtaining a working, integrated, complete picture of an internal cognitive architecture, and of its components and functions (ie decision and reaction, inference and learning, emotions and 'rational' reasoning);
b) completing the coupling between the agent and its environment in order to have an effective, complete loop: perception, cognitive processing, (re)action, effects or independent changes in the word, perception of changes, and so on.
Which of those paths is more important, or should be given precedence? We believe that they should advance in parallel - as is in fact happening - and preferably interact as, on the one hand, a more effective coupling with a complex environment may well require more complicated cognitive architectures and richer cognitive functions, while, on the other - cognitive components and architectures must be functionally justified by showing their advantage and adaptive value when interacting with a constraining environment.
This perception also clearly emerges from the contributions in this issue, which range from neural nets to logic for action, from reactivity to planning, from situated to distributed approaches; and, in some cases, explicitly discuss (although briefly) the issue of anti-representationalism. Two very interesting and converging movements can be witnessed: on the one hand, some research groups are going from sensor-motor, procedural, and situated levels of cognition towards 'more abstract representations'; on the other, there is a trend moving from merely symbolic reasoning towards more grounded, embodied and sensor-motor representations. We can observe a remarkable and explicitly interdisciplinary effort, involving psychology, neuroscience, and philosophy among others.
Thus the picture of Cognitive Systems - as emerges here - not only has various focuses (mainly: mind-body integration in cognitive architecture; multi-agent environmental co-ordination) but clearly shows the fragmentary and incomplete status of this domain. As yet, we have no complete, mature or fully integrated architecture or model; we continue to have several 'isolated components'; exploration for new models is enthusiastic but sometime a bit naive; there is a strong, useful, but premature desire to model technologies for interesting applications. In a word, this is an exciting, promising, growing domain, that will become mature during and (also) thanks to European initiative in the Sixth Framework Programme.
Cristiano Castelfranchi, ISTC-CNR