A few characteristics of the work system designer

Some other studies indeed reveal useful findings which can be exploited in the work system design for the sake of work performance.

• to take the capabilities and the pertinent functionality of human cognition as a performance reference, but not necessarily as an implementation model to be copied,

• to effectively design for the real needs of the human operator in the work process based on the knowledge about the strengths and weaknesses in human cognition,

• to generate models of human behaviour which can be incorporated in artificial components as knowledge components to enable human-like capabilities of effective co-operation with the human operator, and

• in summary, to design work systems by use of artificial cognition on a more systematic way. Some of the findings have already been mentioned right along with the description of the particular implementation principles.

The conclusions in this article, though, do not dwell now on others for the sake of completeness. Instead, they highlight just a few findings, no matter whether they are already mentioned or not, which appear to us as those of most potential effect on the work system design. The vast amount of accumulated valuable knowledge about human factors in addition to that cannot be further commented here anyway, although it is strongly recommended to study these findings, too, if they are not kown yet. Here, we only focus on the following six conclusions:

(1) The finding of probably best yield for the designer of artificial cognitive systems is that about the the two modes of human information processing with one of them being

• standing for the widely distributed unconscious specialised processing subsystems which are essentially working in parallel, and the other one being

• standing for a process which is working on the basis of sequential processing steps (accompanied by conscious experiences) for the purpose to selectively (goal-relevant) drawing information from widely distributed unconscious specialised processing subsystems it can communicate with by means of the so-called global workspace and to integrate this information and to broadcast results among the unconscious specialised processing subsystems.

The unconscious cognition in processing subsystems, which is often called the automatic one, is extremely efficient, if routine behaviour with quick reactions is demanded. It is taking advantage of massive parallel processing. This relates to the by far largest part of cognitive processing activity going on in our brain at all times. The conscious cognitive processing, which is often called the controlled one, is time-consuming but very effective, too, and the only effective one, if higher level behaviour of deliberations is demanded which cannot be provided by the automatic one. As a consequence, this article is proposing that artificial software agents should be furnished with hybrid architectures of that kind layering two kinds of agents, reactive subsystems forming a basic layer for the automatic cognitive processing and deliberative agents in a layer on top for the controlled cognitive processing. The latter may also include agents for meta-management, providing self-assessing, selfmonitoring and self-modifying capabilities. According to these two modes of information processing human cognition features two main principles of memory implementation, explicit and implicit memory. Unconscious processing for routine behaviour takes advantage of implicit memory. Deliberative behaviour, though, which copes with situations demanding for conscious decisions and more effortful deliberations, is characterised by making use of explicit memory.

(2) The content of the explicit memory is semantically coded, i.e. by explicit meaning. These semantically coded explicit memory items (so-called chunks) are represented sub-symbolically, possibly including an abundance of implicitly represented features, which offers a representation much richer as other representations which consist of a limited set of explicit features only. Moreover, semantic coding does not exclude the representation of explicit features either which provides our capability to deal with symbols, too. We might also realise, perhaps as a surprise, that many features, which are also part of the implicitly represented ones of a content item of explicit memory, can be selectively evoked as explicit ones by the global workspace. Thus, human cognition demonstrates how to take advantage of the benefits of both connectionism and symbolism. This kind of knowledge representation

• avoids ontology problems (ambiguities etc.),

• ensures a fine-grained representation in the feature space along with outstanding generalisation capabilities, and

• enables coexisting knowledge representations of almost an arbitrary number domains, independent and dependent ones.

As we think, this is not being sufficiently appreciated by the community of designers of artificial cognition, so far. Unfortunately, the corresponding mechanisms in human cognition are not yet fully explored, but the theory of global workspace as a well-accepted hypothesis might yield a valid basis to generate ideas for how to implement a similar approach in artificial cognition. There is very little undertaking to our knowledge, so far, like that of them for instance, which seriously is pursueing this aspect, but we are sure this is the way to go for future designs. It is also of great interest to the designer of cognitive systems how learning is integrated in the overall system architecture of human cognition by connecting perception with learning, where attention control in the context of the issue of relevance of environmental stimuli plays a role, too. The way of achieving tremendous generalisation capability and equally high reliability as accomplished for human perception has to be taken into account in that context. This is directly pointing to the so-called plasticity versus stability dilemma.

(3)The functional component of learning was neglected in most cases of existing artificial cognitive systems, because knowledge can principally be implemented in artificial systems as a given thing, which might have been acquired offline by knowledge acquisition methods. This admittedly pragmatic approach of system implementation will not necessarily be a leading one for future applications, because online learning will be a valuable design feature for the performance of artificial systems, too, when online adaptation capability plays an important role.

(4) The processes of human voluntary action are pivotal and very complex. The design of these processes indeed account for control dilemmas in a satisfying way like the already mentioned plasticity versus stability dilemma, and in addition the

• maintenance versus switching dilemma and

• selection versus monitoring dilemma.

This addresses the following: Although human behaviour is goal-driven in the first place instead of being data-driven, persistence in the pursuit of long-term goals is demanded and at the same time enough flexibility is demanded to interrupt this pursuit in favour of new dispositions which should be immediately carried out in response to certain urging changes of the situation. In addition, the selection of goal-relevant information has to be designed that it does not suppress information, which is relevant to another competitive goal of higher level. To design for these kinds of dilemmas is a great challenge for the designer of artificial cognitive systems. Further findings in neuropsychological investigations on these processes can be expected in the coming years which might be helpful. These should be followed up as well as findings on the other mental processes, in order to be able to make use of them.

(5) It should be realised by work system designers that the amount of findings about human cognition has been developed to an extent that qualitative and quantitative models of human behaviour are no longer something one only dreams of. Modelling of about all high-level functions concerning human situationdependent action generation based on explicit and implicit knowledge can be accomplished today. Some of the prototype systems of artificial cognition in work systems are significantly taking advantage by making use of these kinds of models.

(6) Finally, some words about limitations of human cognition. An often neglected design recommendation for work systems is the rule to bring the human operators in full action when there are tasks which match ideally with their excellences, and to alleviate him from tasks or support him effectively where his weaknesses could play a shabby trick.

For instance, if visual perception as such is a central task, there is still no artificial means which generally can do better, but if this is coupled with specific demands on attention, for instance, compliance with these demands is not warranted. Here, for instance, support by means of artificial systems can be very effective. Artificial cognition is always vigilant and attentive. There is also no fatigue problem. Also consciousness is not an issue for the design of artificial cognition, because everything which goes along with consciousness like working memory and attention control can be warranted by design. These are typical properties of artificial cognition which can help to avoid violations of human limitations in the work system design. In artificial cognition motivational contexts are also governing the cognitive behaviour, but usually it is not wanted to design a system with the kind of emotions which are insufficiently controlled. The motivational contexts in artificial cognition are defined on a rational basis. However, more than one level of motivational contexts as it is realised with human cognition by the level of limbic motivational contexts and the cortical ones (orbitofrontal cortex) is also for artificial cognition an interesting approach.

A few characteristics of the work system designer

Legal Disclaimer

Related Articles