This monograph was originally written in 1993, to bring together my ideas on a proposed cognitive processing mechanism. In 1994 more material was added on the evidence for the mechanism, and the paper was divided into three independent parts. This had the disadvantage that the close integration of themes - the cognitive mechanism, the data from which it was developed, and the wider implications of the knowledge usage - was lost. In this version all sections have been integrated again. The paper is presented in parts, for downloading and ease of access, but the parts are not written to be read independently.

November 1993, revised August 1994, November 1997

The key features of this account are cognitive goals/needs, a working storage overview, and meta-knowledge. A processing element consists of a cognitive goal/need, with working storage for what meets the need, and links to methods for meeting the need. Behaviour is active and goal/need oriented, rather than occurring only in reaction to environmental events. 'Routines', which are independent from the cognitive needs which they meet, are themselves built from elements. 'Routines' have associated meta-knowledge about their properties. This meta-knowledge is involved in the choice of working method for meeting a cognitive need, and in the control of mental workload and learning. 'Sequencers', consisting mainly of conditional elements, determine the sequence in which the main cognitive needs are thought about. The working storage associated with the main cognitive needs in the 'sequencers' is maintained relatively continuously, and acts as an overview of the current state of the task. This overview is the context determining the optimum sequence of, and method used in, thinking. Serial behaviour, and different modes of processing, emerge from the contextual mechanism in operation in particular contexts, rather than being formally defined in advance.

1. Introduction

Cognitive behaviour is often complex. I find it fascinating to ask whether there is some relatively simple mechanism from which this complex behaviour could be built up, to give a parsimonious account of its flexibility.

The aim of this paper is to describe such a mechanism, called a 'cognitive processing element', and to explore how much of cognitive behaviour it can account for in complex dynamic tasks such as industrial process operation or air-traffic control. Extensive analyses of behaviour in these tasks suggests that their key features are as shown in Table 1. For each of the key task features, the Table suggests the cognitive processes involved and the underlying mechanisms needed. The aim of this paper is to account for many of these features of behaviour. This paper gives some of the evidence for these features, more evidence can be found in Bainbridge, 1981, 1992, 1993.

Table 1 : Key Features of Complex Dynamic Tasks

Features :	Processes implied :	Foundations :
1. Information about state of device, or effect of action is incomplete or ambiguous.	Understanding and evaluation require : a. inferences about past, present, future. b. active search for information. c. 'risky' decision making - expectations and costs affect attention, identification, interpretation, and the cognitive activities below.	i. knowledge as a basis for: inferences, information required, probabilities, costs. ii. temporary inference structure (overview) built up in working storage. iii. this overview acts as a continuing context for other cognitive activities. iv. ability to build up overview develops with experience.
2. Operation or control of external dynamic entities which evolve over time.	a. actions of corrects size and timing. b. anticipatory actions may be more efficient. c. may not be time to consider alternatives.	i. experience needed of control. ii. anticipation involves inference. iii. knowledge enables recognition of situation categories and causes.
3. Situations and/or responses not all pre specifiable in detail.	a. adaptable, flexible, e.g. revise method or goals if necessary b. may need to formulate goals, sub-goals. c. review actions available. d. devise new working methods. e. general strategies needed.	i. non-rigid working methods. ii. goal orientation. iii. goal-means independence. iv. knowledge of general methods. v. actions available are included in overview.
4. If: several control targets, and/or hierarchy of sub tasks, and/or interdependencies in device.	a. allocate mental effort between tasks. b. interleave tasks (multi tasking). c. not necessarily a set sequence of thinking about task topics. d. must cope with interactions between sub tasks. e. plan (consider alternative solutions).	i. maintain overview in working storage, of state of tasks. ii. choice mechanism needed (not necessarily conscious). iii. mentally try out alternative, in working storage.

In my view, the nature of behaviour in complex dynamic tasks means that the focus in accounting for it needs to be on the underlying organisation of behaviour, and on how the behaviour can be flexible in dealing with situations which change in detail from one instance to another. The focus here is also on cognitive processes in well practised cognitive skill. Learning is mentioned in Section 7, but the mechanisms proposed here are not sufficient to deal with all the processes involved in learning.

The most frequently used element in representing cognitive processes is an (if condition - then action) pair. When combined together, such elements often only make up undifferentiated
chains or networks, without any further structure. They may be sufficient to account for simple laboratory tasks and intellectual games, but they are inadequate to account for the behaviour indicated in Table 1. To account for cognitive behaviour in tasks such as industrial process control or air-traffic control, it is not necessarily adequate to expand a cognitive model used to account for simple tasks. A simple model may not contain the concepts sufficient either for understanding the behaviour or to suggest human factors/ ergonomic designs which would support this behaviour.

This paper proposes that many of the underlying mechanisms can be formed from a richer but still simple cognitive processing element. This cognitive element was first applied, and has been used most extensively, in a simulation of the cognitive processes of an experienced steelworks furnace power supply operator. (The task is described in Bainbridge et al, 1968.) This paper will also discuss other examples of complex human behaviour which need to be accounted for, especially when they provide a briefer example, or when they suggest modelling needs which were not required to account for the furnace operator data. The furnace operator gave a running commentary about what he was doing, a verbal protocol. Analysis of this protocol indicated what was needed to model his cognitive behaviour. As the main data analysed came from verbal protocols, so the analysis has all the limitations of this type of data. This problem, and the analytic techniques used, are discussed in more detail in Bainbridge (1985, 1990). The analysis of the verbal protocol was built up into a paper-and-pencil simulation of the operator's cognitive behaviour (presented in full in Bainbridge, 1972, Figures 7.2.1 to 7.3.3). Paper-and-pencil simulation was used, because I did not want, when thinking about the types of mechanism underlying cognitive behaviour, to be constrained to the facilities available in any particular programming language. The emphasis is on the types of mechanism needed, rather than on a specific mode of realising them.

This paper also makes some points about additional mechanisms which are needed to account for the evidence on human cognition, but which might not otherwise be needed as part of a general cognitive device. The paper finally discusses extensions and expansions to the model : how its knowledge mechanisms could account for aspects of behaviour which this approach to modelling was not originally devised for, particularly in the control of mental workload and learning, and some features of behaviour which it does not account for and for which additional mechanisms are needed.

In this account, cognitive processes are built up from processing elements. A processing element consists of a cognitive goal/ need, with working storage for what meets the need, and links to methods for meeting the need. 'Routines', for meeting the main cognitive needs, are themselves built from elements. 'Routines' have associated meta-knowledge about their properties. This meta-knowledge is involved in the choice of working method for meeting a cognitive need, and in the control of mental workload and learning. 'Sequencers', consisting mainly of conditional elements, determine the sequence in which the main cognitive needs are thought about. The working storage associated with the main cognitive needs in the 'sequencers' is maintained relatively continuously. This acts as an overview of the current state of the task. This overview acts as the context for choosing the optimum sequence of, and method used in, thinking (see Bainbridge, 1975, 1978).

In this approach to modelling there are three key theoretical concepts :
complexity : the processes underlying the contextual mechanism are themselves built up from a basic cognitive processing element with potential for adaptability, i.e. complexity emerges from simplicity.
context : complex behaviour involves building up an overview of understanding of the task situation and what to do about it. This overview acts as the context for doing the behaviour, for choosing what behaviour to do next, and for choosing how to do the behaviour. The behaviour done updates the overview, which then determines the next best behaviour, and so on.
meta-knowledge : working methods are chosen on the basis of meta-knowledge about their properties.

The word 'complex' has already been used with two meanings :
1. as in complex tasks, in which the person may need to integrate simultaneous responsibilities for several activities, each of which may not be simple. This use of the word is implicitly defined by Table 1,
2. as in complexity theories, in which complex behaviour and its organisation emerge from simple components.
The aim of this paper is related to both meanings : to account for human behaviour in complex tasks in a way that has the characteristics of a complexity theory. However, I hope it will be clear that everything said in this paper results from data analyses rather than being driven by commitment to a particular theoretical principle. The mechanisms which will be described were first used simply as a notational convenience for representing the furnace operator's behaviour. They were only reified into basic mechanisms when I found that they gave a parsimonious and internally consistent account of the behaviour, and could also answer questions which they had not originally been devised to account for. The length of this paper is due, not to the difficulty of explaining the proposed mechanism, which is quite simple, but of describing the behaviour it is supposed to account for, and how it accounts for it. Hopefully the summaries will help in differentiating the suggested mechanism, which they concentrate on, from the data.

In my previous papers, the focus has been on accounting for, or designing to support, cognitive behaviour in complex dynamic tasks, and those accounts made use of assumptions about the nature of the underlying cognitive mechanisms which were based on my conclusions from the furnace operator simulation. The aim of the present paper is the inverse, to make the nature and potential of the proposed mechanisms as clear as possible.

This paper describes the reasons for suggesting these mechanisms, the detailed nature of the proposed element and the structures built up from it, and the aspects of behaviour which it is claimed can be accounted for by it, in the following sections :
* the basic element, with simple examples of 'routines' built up from these elements, and how the 'routine' uses local working storage (Section 2).
* ways of meeting the cognitive need expressed in a processing element, by data acquisition or by 'routines', how 'routines' are built up, and the main 'routines' found in process control tasks (Section 3).
* 'sequencers', which determine the order in which cognitive needs are thought about, and the working storage overview which they build up (Section 4).
* more on working storage, its structure and capacity (Section 5).
* how the choice is made between alternative methods of meeting a cognitive need, on the basis of meta-knowledge about the methods, and the place of this choice mechanism in the control of mental workload (Section 6)
* the part played by the choice mechanism in learning and in the use of different modes of processing. This paper will raise some of the issues but will not give a complete account, either of problems in learning which need to be accounted for, or of a mechanism which could explain all the processes in learning (Section 7).
* a brief discussion of the nature of the knowledge bases referred to by the cognitive elements during processing (Section 8).
* conclusions (Section 9).

©1997 Lisanne Bainbridge