Skip to main content

Article Category

News

Article available in the folowing languages:

Launch of a New Cognitive Systems Project: COSPAL - Cognitive Systems using Perception-Action Learning

COSPAL's objective is to investigate design principles and architectures for technical systems, which are capable of learning basic cognitive skills in a similar way as humans do.

It might appear surprising that, for instance, learning basic motor skills as done by children during the first three years of their life are hard to simulate in a technical system. To understand why the 'simpler' capabilities of humans are more difficult to simulate than 'higher level' capabilities like, for instance, playing chess, one has to consider the different approaches in human learning and in implementing technical systems, which appear to be intelligent.,Nowadays technical systems typically consist of a predefined set of rules, which control the system's actions depending on the inputs. Unforeseen inputs and constellations cannot be processed by such a system. Playing chess is a very well suited problem for this kind of system design, because the chess game follows exact rules and has a finite (although large) number of distinct constellations. Simulating certain cognitive capabilities of a one year old child, however, cannot be implemented by a rule based system for two reasons: First, there are no exact rules in basic behavioural schemes, they remain too fuzzy and qualitative to be used in a technical system. Second, the number of possible inputs and the number of possible constellations are infinite in a way that for any number of rules there exist cases that are not covered by the rules. Due to these limitations of rule-based systems, several attempts have been made in the past to design systems, which also show fuzzy behaviour and where rules were replaced with different ways of machine learning. In machine learning the technical system is supposed to adapt to a large set of examples, the training data, and shall generalize from the number of used examples to all potential constellations. The new aspects to be considered in the context of the COSPAL project are a suitable learning strategy and system architecture for simple perceptual schemes, where the associations of input and behaviour are triggered by the system's own actions. Furthermore, a suitable interface to rule-based systems is required in order to make the basic capabilities available to a system in a larger context and to solve tasks. In order to evaluate the progress of the COSPAL architecture, a technical system shall do the same what small children do: learn to manipulate simple objects like building blocks and to puzzle objects into a shape sorter.For further details see http://www.cospal.org/ or contact the project,coordinator Michael Felsberg of Linkoping University (E-mail: mfe@isy.liu.se)