Book Review - Human Problem Solving
Human Problem Solving. By A. NEWELL and H. A. SIMON. (Englewood Cliffs, N.J.: Prentice-Hall, 1972.) [pp. xiv+920.] £8·50.
Newell and Simon claim to have made” the first explicit and deliberate exposition of the position now known as ‘information processing psychology’ “ (p. 885), in a paper published in 1958. They have been the first to clarify some of the fundamental issues in this way of theorising about cognitive processes, and their contribution is both influential and substantial. This book provides a readily accessible source on all their work, some of which has previously been available only in privately circulated reports, and it is at first sight both monumental and overawing.
” The aim of this book is to advance our understanding of how humans think “ (p. I), by characterising the human being as an information processing system (IPS). The first chapters introduce the authors’ values and prejudices in developing IPS theories of human behaviour, and discuss the notions of an IPS, the task environ- ment, and problem solving. Major sections present data from three problem solving tasks: crypt-arithmetic, logic and chess. Nearly a third of the book consists of verbal protocol transcriptions from these tasks, which are invaluable both for the sceptic, as there are so far no accepted techniques for analysing this type of material, and for the expert interested in details of the behaviour. A final chapter summarises the theory.
Thc authors introduce the general notion of an IPS in detail, as a background to demonstrating that a human being doing these tasks is an IPS. An IPS consists of receptors, effectors, memory and processor. The memory contains ‘symbol structures’, i.e. sets of related tokens, which can designate either objects or infor- mation processes (programs); these processes can accept one symbol structure as input and produce another as output. The processor consists of elementary infor- mation processes (the basic mechanisms from which the main processes are built up), a short-term-memory (STM) which stores the inputs and outputs of the processes, arid an ‘ interpreter’ which determines the sequence of processes to be used, according to the contents of the STM. Such a general definition avoids equating digital com- puting with human thought, by describing the general class of system of which brain and computer are particular instances.
Their general theory of problem solving involves two basic notions, the problem space and the production system. The problem space is the representation of the task inside the problem solver. It consists (p. 810) of elements; i.e. the possible task components such as the pieces in chess, operators such as the permitted moves in chess, and the knowledge available to the problem solver at a particular time, such as the present’ state’ or position on the chess board and how it was arrived at. The problem space is therefore a function of the task environment, which defines the elements and operators, and the capacity of the IPS’s STM which limits the current knowledge.
To find the solution, the problem solver has to follow a path through the problem space from one possible state to another. Any simple method involves searching a literally astronomic number of paths, so heuristics, which meaningfully limit the search, are used. Newell and Simon suggest that, because of the limited capacity of human STM, these heuristics are carried out by a production system. A production system (a strategy/method/program) defines which operator to apply in which cir- cumstances (present knowledge state), so it generates the sequence. of behaviour. As the production system can only react to and lead to elements and operators which are in the problem space its activities are a function of the problem space.
Unfortunately, although the authors consider their account of an IPS is informal, it is expressed in a parsimonious language of definitions and postulates which may be difficult for the reader with no knowledge of computers to follow. Many of the specific analyses in special notation can be omitted by the general reader, but these definitions are essential to any understanding of the book. It might have been accessible to a wider audience if an even more informal starting point had been used.
Given the book format, Newell and Simon have been able to present their data and analyses in more expansive detail than in previous reports, and the data sections contain many interesting theoretical and practical discussions. Unfortunately, the authors have taken too uncritical an advantage of this freedom. Points are made as they arise, so that generalisations appear in the middle of data, and method and results are not kept distinct. Thus, the reader who does not want to miss any of the generalisations has to skip.read all the data sections; this is both boring and confusing.
It is difficult to grasp the concepts presented in the book, partly because they arc difficult, but partly because of the confusing style. It is also because the authors have two theoretical standpoints which they do not think are distinct but which actually have very different implications. This provides the substance of the criticism which follows.
The general impression is that Newell and Simon are interested in human beings as a source of data on problem solving rather than vice versa, as they consider trivial many factors which one would want to know about the human IPS. Through. out the book they make two types of statement, for example:
The (‘ environment.centric ‘) view simply places the constraints upon the possible forms of behaviour in the environment rather than a.ttributing these constraints directly to psychological mechanisms. It proposes limits on knowledge and how to obtain it, rather than limits on the ability to perform according to knowledge that has been assimilated “. (p. 866)
” the apparently complex behaviour of the information processing system in a.given environ- ment LS produced by the interaction of the demands of that environment with a few basic para- meters of the system, particularly characteristics of its memories” (p. SiG).
Newell and Simon essentially work within the’ environment-centric’ view, although they do mention the other. Several of their assumptions support this view, in particular that the problem space is ‘ demanded’ by, and a simple mapping of, the environment. They apparently consider that program and long-term-memory (LTM) organisation are insignificant, and that a simple state/process account of the problem space is sufficient. These can be criticised in turn. One can argue that psychological mechanisms have a fundamental influence on behaviour, and one can suggest that a sufficient theory of the human IPS will differ from the one developed .here. Their problem solver may be able to manage logic, but it could hardly cope with Sloman’s” spilt milk, ungrammatical sentences, unfamiliar typewriters, broken suspenders, lost keys, mixed metaphors and veiled insults”.
Considering problem spacc determined by environment, the authors present ambivalent statements about this, for example:
” the task environment rcmnins tho overwhelming determinant of the problem space.” (p. i90)
” we believe that there arc deep structural relations between the organisation and parameters of tho IPS find tho problem space. respectively-that many of the characteristics of the latter are udnptnt-ions to constraints imposed by the former” (p.SI2).
We can concentrate on showing that the task environment is not the overwhelming determinant of the problem space.
Newell and Simon acknowledge that the environment cannot be completely represented in’ the problem space (Chapter :3), and that there is a difficulty in defining which parts of the environment should appear in the representation of a particular task, They solve this by saying:
“wo mean by the structure of the environment precisely the set of invariants that are preserved under translation from any ouc of these isomorphs to any other.” (p.825).
In other words, the task invariants are those which will appear in any IPS’s internal representation of the task. Thus in practice the invariants, or structure, of the environment can be found only by studying possible problem spaces for doing the task.
This has several effects. It allows the possibility that there are variants in the problem spaces, which may not be a function of the task environment. However, if one concentrates on determining the invariants, as one would within the view that the environment determines the problem space, then these variants are of minor interest and the analysis will not be sensitive to specific characteristics of the human problem space.
The task analysis has to become, as it does for Newell and Simon, a study not of the task environment but of possible problem spaces, which are found by looking at alternative programs or ways of doing the task. As the task invariants can only be found post hoc, determining them has no necessary prior status in analysing a particular problem space. The task environment must of course delimit the meaning- ful ways of doing a task, but the variety of possible programs discussed by Newell and Simon shows that it does not delimit which of these is used by a particular IPS. If this is so then it cannot be the overwhelming determinant of the problem space. Nor can the study of programs for doing the task be necessary, prior or sufficient for identifying the strategy of a particular IPS. This is also easily demonstrated from Newell and Simon’s data. Their subjects show behaviour which has not appeared in any of the programs previously described for doing the task, e.g. the’ progressive deepening’ strategy in chess, or S8 who tried several different problem spaces for crypt-arithmetic. Some of the programs, like GPS, were developed after an analysis of human performance. In chess much of the behaviour shown in common by human and computer players comes from the common chess culture rather than the task requirements.
Thus, these programs show only that the task can be done by some IPS, and provide “ insight into the structure of the environment and the nature of the relevant features in it “ (p. 769). They are neither necessary nor sufficient in doing this for a particular LPS. As a theory, such programs have the limitations of any normative model. The view that: ‘;
” one must work with task environments in which artificial intelligence has provided the re- quisito array of plausible mechanisms” (p. 0).
is unsupported.
The prior status of the task environment in determining the production system and surutegy is also questioned by types of task which Newell and Simon diclnot study, No It priori definition of a task would predict the data-gathering strategies which are a prominent feature of human behaviour in complex environments. In ill- structured tasks where the relevant data is not clearly defined, perhaps methods are not determined by the structure of the problem space but the other way round.
It would only be possible to maintain that the problem space is determined by the environment if one also holds, as Newell and Simon apparently do, that thc problem space is a simple mapping, a straightforward representation, of the environ- ment. Ifthemappingisnotsimplethenthemappingprocesses,whichareproperties of the IPS, are important determinants of the content of the problem space.
Newell and Simon maintain that:
” any structure … in the task environment can be expressed by translating into one of the problem spaces for which its expression is simple. The translation itself may be a complex process, hut it certainly can be carried out by finite means.” {p. 825).
This may be true in theory, but in practice one can ask whether all problem spaces can be represented by a human being. If not, then can a problem space for which the expression of this structure is simple be represented by a human being, and can a human being carry out whatever finite means are necessary to make thc translation? If the answer to any of these questions is no, and one can suggest that a simple mapping is not possible for a large, complex, noisy environment (this is discussed further below), then the human problem space is not necessarily a simple represen- tation of the task environment. Any arguments which depend on the environment- problem space relation may then be questioned, and the nature of the representation and of the mapping processes become basic factors to be studied.
Newell and Simon also say that:
” since the effectiveness of particular heuristics is a function of the problem space, these results .. . do not depend on specific propert-ies of the information processing system that carries out the search. Such generalisations belong to the theory of problem environments rather than to psychology-that is, they are independent of any very specific characteristics of human cognitive processes.” (I’. 139).
This is true only if the problem space is a simple mapping of the environment. If the problem space does depend on specific characteristics and Iimitations of the human cognitive processes, which construct and use the internal representation, then the effect of the heuristics will do so as well.
Considering program and LTM organisation, the authors say:
” the question of program organisation is as irrelevant to problem solving as the question of perceptual organisation. For all programming languages have about the same power of expression -anything that can be stated in one programming language can be stated in the others.II (p. 803).
If one holds this view and is interested in problem solving then details of the human information processing language may well be unimportant. On the other hand, if one is interested in the nature of the human IPS, even if this statement is true it is irrelevant, as the interest lies in finding the nature of the human programming language.
In theory, It universal Turing machine can model the behaviour of any other computer. In practice, the translation from one machine to another may pose non-trivial problems. If some expressions are more compatible with the human IPS than others, then one would like to know which. Nor does the human IPS have unlimited storage capacity and time. Consequently the organisation and language of the human program may not be irrelevant.
” one can predict the problem solving program ofan IPS only after characterising the (memories) av-ailable to it..” (I’. 803).
In part Newell and Simon’s characterisation is:
“The human memory is usually described as associative. Associativity is achieved … by storing information … in symbol structures, each consisting of a set of symbols connected by relations. . . . The kind of associativity described … seems appropriate for the tasks we have examined. . .. But our tasks do not involve extensive processing in long-term memory, which might call for other forms of eeeociatdon.” (p. 792).
In a sense any memory is associative, as it must consist of linked items. Newell and Simon seem to mean simple associativity, in which case theywould consider the organisation of LTM straightforward. They discuss only timing aspects of the storage, retrieval and accumulation of knowledge, and there is now reference to mcmory organisation.
Their data may only require simple associativity. Other forms of LTM may exist however: it may be content-addressable, or the symbol structures may be programs (which compute the data required) rather than representations of the data. If L1’M does contain forms of associativity which are necessary in other tasks but not here, it may be that they arc used here although they have greater power than is required. If this is so, then any notions based on an assumption of simple associa- tivity do not necessarily hold. Such notions are those of simple mapping, discussed above, or that learning is uninfluential and perceptual processes irrelevant. Their discussion of these factors further supports the interpretation that they consider only simple associativity.
In their view:
” experience is simply one more source of information that can be exploited to attain adaptive performance.” (p. 137).
” The table of connections . . . (which specifies which operator is relevant to reduce a particular difference) … can be found simply by matching the output expression of the operator to the input.” (p. 435).
This is a good redescription of conditioning-type learning theory, with simple re- lationships between input and output and also between external environment and internal representation. Instead one can suggest that in most tasks the en- vironment, or the effect of an action, is so complex and variable that extracting its statistical properties is insufficient to make sense of it, and some pattern induction and recognition mechanisms are required. Then experience is not.’ simple’ at all, but it provides new data which is perceived and understood in terms of the present structure of the internal representation of the environment. The new data may also lead to restructuring of the internal representation. These processes are germane to the nature and functioning of the problem space and LTM.
Somewhat similar comment can be made about their views on perception. In discussing the interrelation of perception and problem solving in the context of chess (p. 775), they argue that whole groupings of chess pieces are recognised as units by expert players-‘ chunked’ together in Miller’s terminology. These chunks have automatic overlearned responses which do not require present problem solving. There is no discussion of the nature or origin of these chunks, and their arguments imply that chunking processes are independent of, and irrelevant to, problem solving. However, it seems likely that the recognition of groups is related ~o the internal representation of task knowledge. Thus, problem solving and chiinking may interact, or at least findings from the study of one may have implications for the other.
Newell and Simon do mention the possibility of more complex processes in LTM:
” There is an intriguing possibility that a. production system offers a viable model of LTM. Possibly thoro is no LTM for facts distinct from the production system-that is. no basic distinction between data and program; rather the LT~I is just a very large production system.” {p. 805).
” all increments of knowledge are cast in the form of productions” (p. 867).
The production system is a program which decides, on the basis of present knowledge, what to do next. There is no further discussion of the point quoted, but if it is to be more than it tautology then one wants to know more about how it will work. Ques- tions about organisation and structuring in the problem space now become questions about the production system. In particular, if data and program are not distinct, this has strong implications for the relation between problem space and production system; Newell and Simon maintain these separate throughout.
Although Newell and Simon have not discussed the organisation and processing determining the contents of the problem space, they do indicate different types of processing operator used in doing a task. The points made so far about the deter- minants of the problem space apply even if it contains only descriptors of thc environ- ment. If there are elements and operators used in doing the task which do not describe the. external environment then these must be a function of the IPS not the environment’, and so further question the usefulness of that distinction and indicate that the mapping cannot be simple. Such elements and operators do appear:
” In describing the logic task initially … we took the element of the problem space to be the logic expression and the operators to be twelve rules-essentially the basic space as defined by the experimenter … we modified this specification by taking as the operators the goals, rather than the rules. The ourronb goal provides a subst.antdal amount of the information used in select- ing the next operator, so that the adoption of a.new goal (such as to apply an operator) changes the state of knowledge substantially. Therefore it must be included as part of the problem space element.” (p. 580).
It seems that whenever new elements and operators are found these are put in the problem space by definition, without questioning whether they have a different status. Productive and attentional operators (p. 844) and search and inference operators (p. 583) are distinguished, also different search strategies may usc different operators and so imply different problem spaces (p. 827). The problem space includes elements for describing the present knowledge state and the route by which it was reached (p. 810). These different types of element and operator are associated with different types of production, such as recognising a solution, evaluating a present state, or selecting an operator (p. 837). The names of these operators and productions indicate that they not only describe and manipulate the external world but also the internal representation of the world. These different operators and productions are mentioned piecemeal, with no overall integrated discussion. If one wants to know the range of possible human information processing behaviour then an overview of the types of element and operator which are found would be both interesting and important.
This variety of elements and operators indicates the simplicity of the statement:
” selecting an operator … involves functions from states to operators, connecting the language in which states of the world are described with the language in which processes that act on tho world are described-the afferent language with the efferent.” (p. 831).
‘Afferent’ and ‘efferent’ usually apply to interactions with the external world. The types of operators found show that they must at least also apply to manipulations of the internal representation of that world. The statement is most obviously inadequate when one considers the necessary components of an internal representation for a dynamic autonomous external world. Not all changes in this external world occur because of actions by the IPS, so the internal representation of it must include a description of its own operations. Even though all these different types of operation may be expressed in the same language this indicates that any simple state/process distinction is inadequate.
The notions that mapping is simple and different types of operation are unimportant lead to a simple account of planning behaviour.
” Problem solving in the context of a ,plan has somewhat the same flavour as executing a. specialised algorithm, for plans are structurally identical to programs. They are symbolic structures, available in LTM … , that are used to guide act.ion in exploring the problem space.” (p.822).
This minimises the importance of planning, particularly if it is an algorithm rather than the more sophisticated heuristic, which is used in problem solving. The state- ment is not true if the plans are not already in existence. Learning about the task environment has already been discussed, but we can comment here about learning (or developing, planning) new task strategies.
Newell and Simon do distinguish (p. 840) between subjects who have assimilated the task rules and simply perform them (the operation is available directly as a production system) and subjects who apply the rules from the rule shect (the’ opera- tor’ on the rule sheet is now data which a production system’ interprets’ to ensure that it is carried out).
” Though this book is not tho place to elaborate this point ((why not?}), the distinction between ussimilatdon and interpretation provides a structural basis for distinguishing some varieties of meaningful versus rote . . . learning.” (p. 843).
Newell and Simon see this distinction as a subsidiary problem of effecting a method, independent of its main function (p. 845). If one is interested in human beings rather than problem solving this distinction is certainly not trivial, as interpretation is an intermediate stage in developing new production systems.
Someone learning to solve a particular type of problem has two levels of goal, to find the solution to this particular problem, and to find a general method for solving this type of problem. The concepts of state of knowledge, operator and problem sIJacO are relevant at both levels. Finding a method may require various specialised operators and evaluation functions.
I was interested in the claim (p. 203) that 80% of the sequencing in a crypt- arithmetic protocol was accounted for by the program, as 80% of my protocol data (from a very different task) is accounted for by a similar type of program eon- taining procedures and sequencing determinants (derived in a different way). The nature of the variability in the remaining 20% suggests that the routine 80% des- cribes overlearned experienced behaviour, and that the subject has the flexibility to generate new sequences reflecting his overall knowledge of the task. It suggests that the ultimate account of task knowledge will be in a form which can be accessed in different ways for different purposes, with the routine behaviour embedded as an over-learned special case.
The lack of an overview of the operator types found by Newell and Simon is paralleled by a lack of general comparison between the different data analysis tech- niques used. What, for instance, are the implications of the fact that different problem-solving tasks require slightly different analytic techniques! In what ways might these techniques be inadequate as a total set of techniques to analyse all cognitive tasks! Might such a set of techniques give an operational definition ‘of the different types of mental task!
Newell and Simon’s restricted attitudes may be a function of the particular task types they have studied. These tasks involve a static closed external world, there usually can be a simple mapping from external to internal elements and operators, and little perception, memory or learning is involved. Concentrated attention on such tasks may not raise the issues discussed in this criticism.
Claiming that the theory (solution) they have found is a function of the environ- ment in which they worked might be taken as an example which supports their own theory, but it is the end product which has been determined by the environment, not the processes by which this was found. These are more obviously a function of their IPSs, particularly their background in artificial intelligence! There are certain basics to their theory which cannot be gainsaid, such as the production system and the notion of means-end analysis, but these can be considered the task invariants; as they would be induced by anyone sufficiently creative and intelligent who studied this type of data. Newell and Simon must of course be congratulated and respected for being the first to do this. Given these invariants, the mechanisms by which they are implemented may ultimately turn out to differ from .the ones presented here. Apart from an explicit account of the different functions and levels of operation in the human IPS these may include a general data base organised so that it can be accessed for different purposes and containing pattern recognition and induction processes.
Many readers may think that these criticisms are not severe enough, as they all accept the framework within which Newell and Simon work, the assumptions that verbal report data and program simulation are useful tools for studying cognitive processes, and that these involve internal representations. There are other objections, that Newell and Simon study abstract symbolic tasks which are inherently verbalis- able and do’ not involve imagery or aspects of problem solving, or that they have’ not tested their hypotheses by experiment, or that the digital computer is an inappro- priate tool (literally or conceptually) for modelling human cognitive behaviour, or that a completely specified theory is inappropriate at this stage of our understanding.
Despite all these criticisms, my reaction to this book is wildly ambivalent, every page containing fascinating insights and statements that I profoundly disagree with. After all, if it was not for Newell and Simon’s pioneering conceptualisations one would not have a framework within which to make these objections.
Although this is a convenient single source of their work it would be tough reading for the beginner, but for the expert the detailed analyses of cognitive tasks will be invaluable, and those interested in this type of cognitive process will have to struggle through this book to work out these issues for themselves.
LISANNE BAINBRIDGE
ChangeLog
- 2018-07-21 李亮创建