On the structure of ill-structured problems: An empirical study on the development of an HIV vaccine

Transcription

1 On the structure of ill-structured problems: An empirical study on the development of an HIV vaccine Paper submitted for DIME Final conference 6-8 April 2011, Maastricht UNU-MERIT & School of Economics and Business, Maastricht University Panel 6. Dynamics of technology and knowledge Pádua, M.(1), Custodio L., (1), Orsenigo L. (2) 1- Instituto Superior Técnico, Portugal 2- University of Bocconi and University of Brescia, Italy Corresponding address: Instituto Superior Técnico Av. Rovisco Pais nº1, Portugal

2 On the structure of ill-structured problems: An empirical study on the development of an HIV vaccine How to accelerate the discovery of an HIV vaccine? To deal with this question, this paper proposes to view the development of an HIV vaccine as an ill-structured problem (ISP) in order to examine the dynamics of knowledge production associated with the search for a solution to this problem. It is argued that the application of such an approach might be useful to deal with the question under analysis, given that problem solving theory has made substantial contributions towards our understanding on the ways through which we can relate processes/heuristics with outcomes (a solution to a problem). Based on the application of such approach, we might be able to identify the patterns of knowledge dynamics that are necessary to reach a solution to this problem and subsequently to gain an understanding about ways to accelerate its discovery. However, even though such an approach might be promising to deal with the question under analysis - its articulation engenders a new set of issues - regarding ISP themselves, which need to be tackled beforehand. Indeed, although the application of problem-solving theory to innovation studies has been helpful to gain an understanding of a variety of problems (ranging from near decomposable problems to complex problems), ISP remain poorly understood. This is due to the fact that the literature on ISPs either tends to ignore non linearity between elements (or complexity), e.g., the fact that the whole is more than the sum of the parts, or, when it deals with this issue, we obtain mixed findings. For this reason, the issue of non linearity requires further theoretical and empirical work. This paper builds on existing theory on ISPs and compares it with the data obtained through this case study in order to generate new theoretical insights regarding the issue of non linearity between elements. Based on these new insights, this study hopes to outline some key features of an approach that can help us accelerate the discovery of an HIV vaccine. The background against which this paper draws upon concerns the ongoing debate about the nature of ISP inter alia the methods required to solve them (the extent to which it requires). With regard to an ISP, three major positions have emerged in the innovation studies literature. An early position introduced by Newell and Simon (1969), Simon (1969, 1973a, b) considers that, well-structured problems (WSP) and ISPs define a continuum and, that ISPs are a sub-class of near-decomposable problems. Another line of thought (Dosi and Egidi, 1991; Dosi et al. 1997; Levinthal, 1997, Pádua 2008 just to name a few) argues that ISPs and complex problems do not define a continuum because ISPs involve a more fundamental lack of knowledge of the problem: the fact that our representation of the problem is not a valid abstraction of the problem (which goes against a linear view of problem-solving). A third line seems to have a position in between position one and two. It argues that ISPs involve some form of local search (Flemming, 2001, Schoenmakers W., Duysters, 2010). In other words, at least some form of linearity is involved in solving ISP from this third perspective. It should be added that, although it is not clear on reading these studies whether other properties are involved, it is stated that at least a local search is involved when dealing with ISPs (in the sense that these studies do not a priori exclude non-linear methods to solve ISPs, as stated by Simon (1973), we have classified them as the third category. This paper proposes a case study methodology to map dynamics of knowledge production for the case study under analysis. It uses co-word analysis to account for the dynamics of knowledge production. Moreover, it incorporates a template or a proxy for

3 a knowledge representation against which it is possible to map these key words and their changes over time. Findings suggest that, alongside Simon, ISP have a structure. There is a hierarchy thereby supporting Simon s insights on properties of ISP problems. Moreover, alongside Kaufman, order and instability co-exist. In other words there is some form of order in rugged landscapes that are near the edge of chaos. Secondly, on a more fine-grained analysis, the study shows that there are several types of ISPs which emerge over time and that the patterns through which these sub-problems evolve and are solved varies according to distinctive parameters. More specifically, it is shown that Simon s views (1973) can account for some parameters that underlie the solving of some ISPs whereas other ISPs cannot be solved in a linear way. In other words, the three positions stated in the literature regarding the structure of ISPs correspond to distinctive types of ISP. Finally, suggestions about ways to organize the HIV vaccine search are proposed. Suggestions go well beyond the so called parallel economics of search (so far one of our bests bets to organize search in conditions of uncertainty) and propose a layered systems view to innovation where we consider different scales of problems solving (from micro to macro level). 1. Why a problem centered approach to the development of an HIV vaccine? Beforehand we need to provide some background information on the development of an HIV vaccine. Firstly, we will define what a vaccine is and then we will put forward arguments supporting the view that the HIV vaccine is the best solution to eradicate AIDS. Secondly, we will identify the nature of the unknowns which surrounds the HIV vaccine and therefore the type of discovery which is required for the creation of an HIV vaccine. It will be shown that the HIV vaccine requires a radical invention. Thirdly we will explain why we think that a problem centered approach to the HIV vaccine can be a promising approach to deal with the question of how to accelerate the discovery of this vaccine. Firstly, we need to define what a vaccine is. A vaccine is defined by Parslow et al. (2001, p. 74) as a substance that teaches the body to recognize and defend itself against bacteria and viruses that cause disease. An AIDS (Acquired Immune Deficiency Syndrome) vaccine is for the time being an hypothetical concept. But it would teach the body to recognize the human immunodeficiency virus (HIV) that causes AIDS and

4 provoke an immune response that would defend against the virus if it entered the body, preparing it to fight and also to remember how to fight, if exposed to this specific infection. A vaccine is not a cure, but prevents infection or slows disease progression. Secondly, we need to introduce some elements of why the HIV vaccine is the best solution to eradicate AIDS. The search for an effective HIV vaccine has become a social and economic priority on the global policy agenda. Its importance is related to its potential role in controlling an HIV/AIDS pandemic whose economic and social repercussions are affecting entire countries and continents. HIV is the leading infectious killer disease worldwide surpassing tuberculosis (TB) and malaria and is the fourth largest cause of death (UNAIDS/WHO, 2006; 2004). According to the UNAIDS/WHO report (2006) more than 33.2 million people have died since the onset of this epidemic and over 40 million people are living with HIV, nearly all of whom will die from AIDSrelated complications by the late 2020s. Moreover, an estimated 5 million people are newly infected with HIV every year. The HIV/AIDS epidemic has reached the industrialized and developing countries but it mostly affects developing countries. In developing countries, the HIV epidemic poses a serious threat to economic development and poverty alleviation, particularly in sub-saharan countries where the epidemic is more advanced (UNAIDS/WHO, 2006). Public health institutions have placed enormous emphasis on the potential of an HIV vaccine to bring the global HIV epidemic under control, for four main reasons. Firstly, vaccines have proven to be among the most powerful and cost-effective disease prevention tools available. Vaccines are considered an important strategy for disease eradication and preventive immunization strategies. An HIV vaccine would therefore be the most cost-effective solution since, although anti-retroviral therapies have extended

5 the lives of patients in industrialized countries, they are out of reach of those living in the poorest developing countries because they are expensive and complex to administer effectively in these settings. Second, anti-retroviral therapies are not a cure and do not eradicate the virus. Third, viral resistance to these drugs is growing. Fourth, the history of public health initiatives suggests that with epidemics of infectious diseases, such as smallpox and polio, only mass immunization with an effective vaccine can bring the required results. However, despite its importance we still do not have an HIV vaccine. While the reasons that are provided to account for the fact that we do not have discovered an HIV vaccine vary - they range from a knowledge gap to a lack of funding (Archibulgi, D., Bizzarri, L., (2005) - it is certainly correct to argue that the scientific and technological uncertainty which surrounds it, is a major obstacle. Indeed, thirty years after the first trials to conduct and HIV vaccine, we are still unclear about the scientific and technological principles guiding this case of product innovation. As radical invention is defined as involving the definition of a new technological principle we will argue that the HIV vaccine requires such type of discovery. In order to deal with this case of radical invention we will build on problem solving theory. This might be a promising approach. Indeed, reflecting the fact that the development of an HIV vaccine is one of the most challenging issues of our times, the sources of its inherent difficulties have been object of major debates but the later tends to be confined to immunological circles (Fauci, A.S., 1996; McMichael A., Hanke, T., 2003) rather than relying on an explicit theory of knowledge production which enables to relate outcomes with solutions. By using problem solving theory we argue that we can gain a better understanding about the major cognitive determinants that make this

6 problem so hard to solve. Namely, we aim to go beyond a descriptive stage (which is used in immunological studies which relies on rather descriptive categories to account for the difficulties associated with the discovery of an HIV vaccine, such as monkey puzzles ) and, instead, offer a description of the complexity of the problem based on a set of parameters. The ultimate goal of the paper is to gain a better understanding of the learning mechanisms necessary to accelerate the discovery of these solutions. This is an important step. If we fail to understand this we run the risk of learning by trial and error, and thereby looking for a solution in infinite time. 2. The debate on how radical inventions come about This section briefly reviews the literature on how radical inventions come about. It focuses on a key issue in this debate which is whether the later results from the recombination of existing knowledge or it results from a radical departure from existing knowledge. This is an old debate that dates back from Schumpeter (1939) and Usher s (1954) pioneering work on the nature of invention in the 30 s. Usher argued that radical invention comes about through acts of insights, e.g., it involves a major mental reorganization. Schumpeter put forward the idea that they result from recombinations of existing knowledge. This debate has been further pursued alongside this main line, with arguments pro-recombination and others pro-major mental reorganizations. While the former supports a linear view of the processes through which radical comes about, the second supports a non linear perspective.

7 In the first section radical invention is defined. In the second section we put forward the argument that radical invention can be viewed as an ISP. Thirdly, a review on the dynamics of knowledge production that take place for radical invention is provided Definition of radical invention Following, Schumpeter s argument that technical change proceeds through the successive phases of invention, innovation and diffusion, scholars have developed theories that characterize these different phases of technical change. Invention corresponds to the creation of new technologies, the innovation phase represents their first commercialization and the diffusion phase concerns their widespread use in an economic context. Regarding the invention phase there has been a controversy about the definition of radical invention. Underlying this debate on the definition of radical invention is the following question: what is novelty (in the context of invention)? The different ways in which this debate can be cast have been identified by Kuznets (1962) who identified two major issues, the definition of novelty, and the order of magnitude of this novelty. The way he framed this debate on novelty is still relevant. For this reason the brief literature review that follows is organized along the major lines proposed by Kuznets (1962). A first line of argument is that novelty can have a broader or narrower focus. It can refer either to the general definition of a combination of existing knowledge in new ways or it can define a more radical departure from what exists. With regard to the order of magnitude of an invention Kuznets (1962) identified three major issues along which the debate can be framed. A first line of argument is that invention may differ in terms of degree of novelty. With this in mind, researchers have argued that (radical) inventions

8 differ in terms of degree of magnitude (even within the narrow view of radical invention we can have different types of radical innovation ranging from the steam engine to poly chain reactions) or, alternatively, they can assume distinct types. Another debate regarding the order of magnitude of an invention concerns whether novelty should be defined from a more internalist or externalist view. Invention can be defined either in terms of its nature (or degree of complexity, an internalist view) or its impact (an externalist view). In this sense people looked at the impact of invention with regard to whether it was economic or technological. A third debate concerned what had been gained in terms of what existed before. This topic is also positioned within an economic or technological perspective. From an economic viewpoint authors look at savings on resources. In technical terms researchers look at the technical problems that have been overcome by the invention. Firstly, Kuznets (1962) among many other economists argued that invention refers to the combination of existing knowledge in new ways. Possibly, this is the most common definition used by economists. Within this very general definition, no distinction is made between radical or minor improvements although it is far the most frequent definition of invention. A few authors have argued that invention implies a more radical departure from what already exists. They define invention as involving a change in technological principle. These authors include Constant, (1974), Vincenti, (1980), Dosi, 1982) and Arthur (2007). Arthur (2007, p. 278) defines invention as a change in a base principle by which a purpose is achieved. Based on this definition, he provides examples of invention like radar technology and others.

9 In reviewing the theories of invention in his 1959 paper, Nelson quotes the work of Rossman (1931) who divided invention into two broad classes: basic inventions and improvements to existing basic equipment/processes. Whereas the former involve applications of new knowledge, improvements are more marginal types of invention which do not involve new elements. Usher also subscribed to this dichotomy and developed it further. He provides a categorization of different types of invention. He distinguishes primary from secondary inventions. The first type refers to a change in the principle guiding an invention whereas the second concerns the practical application of this principle. Usher gives the example of Galileo who was credited with the invention of the pendulum clock. 1 Galileo identified the principles underlying the pendulum clock. By contrast the work of Huygens has been treated as an explicit achievement of the discoveries and inventions of Galileo. Secondly, with regard to the order of magnitude of an invention three major issues have been identified. The first concerns the argument about whether it is possible to define different degrees or orders of magnitude or, alternatively, types of invention. Again the dichotomy between a more radical type of invention and minor improvements emerges. Mokyr (1990) subscribes to the radical view and distinguishes macro-micro- (radical) invention. He defines the latter as "the small, incremental steps that improve, adapt, and 1 Galileo introduced a new principle to account for the motion of objects. He focused on the fall of objects to make his case. As put forward by the website of the Galileo project (assessed in 2008): Whereas in the Aristotelian paradigm physics argued that bodies fell faster than lighter ones due to their weight, Galileo argued that the difference in speed was related to the density of the body. Due to this approach to motion, Galileo was intrigued by the principles which could govern the back and forth motion of a suspended weight. Galileo s discovery in 1602 was that the period of a pendulum is independent of its amplitude the arc of the swing the isochronism of the pendulum. According to the same website, this important discovery had major implications for the measurement of time, namely time intervals. It provided the idea that pendulum could be used to improve clocks. Galileo had this idea in 1641 to use the pendulum as the basis for a more accurate clock but it was Huygens who made the practical application in 1656.

10 streamline existing techniques already in use" (Mokyr, 1990, p. 13), whereas macroinvention emerges as discontinuities from what existed before, such as the steam engine. Typologies of invention were proposed by Stankiewicz (2000). He distinguishes design versus discovery driven innovation. The former depends on an understanding of nature, while the latter is driven by application and demand issues. A second line of debate concerns whether we should adopt a more internalist or externalist view to define invention. This debate reflects the fact about whether we should look not only look at the technology per se but also at its impact (an externalist view). The externalist perspective can be viewed along two dimensions, an economic and technological one. Its economic impact is related to the revenues it generates. Technological impact refers to whether it generates other types of inventions. Possibly the economic impact has been one of the most dominant ways to discuss the order of magnitude of an invention. Conversely, technological impact is more difficult to measure and requires a more qualitative type of analysis. An example that shows how limited are the internalist conceptualisation of invention is the case of the hot air balloon. It was a radical innovation in the sense that it involved a change in technological principles, but both its economic and technological impacts were quite limited. Thirdly, it is possible to discuss the resources that have been saved and the technical problems that have been overcome with the discovery of a new invention. This is a much under-researched area as it requires very detailed knowledge of a technology.

11 For the purpose of this paper we will hold to Constant s (1974), Dosi (1982), Arthur (2007) definition according to which invention involves the definition of a new core or operational principle Radical invention and the dynamics of knowledge production Two major types of positions regarding how radical invention comes about have been put forward. While Schumpeter s proposed the analytical distinction between invention/innovation/diffusion (1939), Usher s (1954) major contribution was to develop a theory of invention. While Usher s contribution to invention encompasses many aspects, we will focus on his views regarding patterns of knowledge production that characterize radical innovation. Usher s contribution (ibid) was a major leap forward, in relation to the theories that were dominant in the time. Firstly, he provided not only a definition and a classification of major and minor types of invention, but he also provided a comprehensive account of how both types of invention are produced. Secondly, like Schumpeter, he viewed technical change as resulting from major breakthroughs which alternate with periods of knowledge accumulation. Moreover, Usher s detailed knowledge of the machine industry helped him to provide a process view of innovation based on Gestalt psychology 2. Based on this view, he proposed a disruptive view of invention. This was an important contribution because the theories of invention prevailing at this time had a cumulative (instead of a disruptive) view of knowledge production (and amongst the advocated for this cumulative view was Schumpeter who argued for a recombination thesis of invention). These two camps or 2 Gestalt psychology is a theory of mind and brain that proposes a holistic view of the brain. The gestalt theory was proposed by Max Wertheimer who was against Wundt s approach to the brain. Wundt argued that consciousness was about the search for basic mental elements. The underlying idea of Wundt s approach is that conscious experience can be broken down into basic conscious elements that can in turn be broken down into basic elements. The Gestalt theory put forward a new principle according to which the whole is different from the sum of its parts.

12 views about how radical invention comes about are still present. Based on these two positions, advocates of a pro-recombination and pro-mental major reorganization have made their way through Major contributions to a theory of invention Usher (1954) reviewed the theories of invention that were dominant at the time and proposed a third one. He identified two existing theories. The first was the transcendentalist theory which he rejected because of its very limited power to explain invention. The essence of the transcendentalist approach is summarized in the following: The transcendentalist approach attributes the emergence of invention to the inspiration of the occasional genius who from time to time achieves a direct knowledge of essential truth through the exercise of intuition (Usher, technical change and capital formation, p. 526, quoted by Ruttan, 1959, p. 600). Usher (1954) criticized this theory in two important ways. Firstly, that this approach neglects the importance of smaller acts of discovery which are also important for invention to occur. Secondly, no explanation was provided for why and how invention occurs over time. Although the transcendentalist approach places these acts of genius in a time context it does not provide any historical analysis of how these inventions are produced over time: The transcendentalist thinks of invention as an array of independent items, each of them of massive proportions, each of them achieved by a process of revelation or intuitive perception that admits no analysis (Usher, 1966, p. 61). The second theory reviewed by Usher was the sociologist one. The major proponents of this view were Gilfillan (1935). They subscribed to the cumulative synthesis hypothesis. Major inventions resulted from small accretions of knowledge. In some

13 ways, the sociologist view of invention takes the opposite position to the transcendentalist school. Whereas the latter argues that there are some kinds of radical insights which lead to invention, for the sociologists there is nothing more than additions to existing knowledge. However, Usher (1954) ended up rejecting both as they adopt an excessively mechanistic view of invention. These two theories (together with the transcendentalist school) regard invention as resulting from a series of cumulative efforts but neither provides any explanation for why and how this process occurs over time. Based on some form of social need or the occurrence of a breakthrough it is assumed that invention will somehow happen in an automatic, predictable and mechanical way. In these theories there is no role for chance or serendipity. For this reason, these are excessively deterministic. Usher then proposed a cumulative synthesis theory which in many ways goes beyond these two sociological perspectives. Usher does not subscribe to a deterministic or mechanistic view of invention. He draws on the insights into the mental processes provided by Gestalt psychology to capture the process of invention. Usher outlines a stage theory to account for individual invention. Four steps are identified. Each step requires an act of insight and he provides examples of the type of insight required for each stage: 1) perception of a problem in which an incomplete or unsatisfactory pattern or method of satisfying a want is perceived 3 ; 2) setting the stage, in which 3 Usher applies this example to the history of the reciprocating steam engine: The emergence of an unsatisfactory need was impossible until there was considerable knowledge of the properties of steam and of the phenomenon of air pressure. It was necessary to distinguish steam from air. It was fundamental to perceive the possibility of producing a vacuum by the condensation of steam in a closed vessel. It was essential to understand the primary principle of atmospheric pressure. Each of these scientific discoveries involved a major break with the traditions of classical and medieval science (Usher, 1966, p. 70).

14 the elements or data necessary for a solution are brought together through some particular configuration of events or thought, among them being an individual who possesses sufficient skill in manipulating the other elements; 3) the act of insight in which the essential solution of the problem is found (Usher stresses the fact that large elements of uncertainty surround the act of insight, and it is this uncertainty which makes it impossible to predict the timing or the precise configuration of a solution in advance); 4) critical revision, in which the newly perceived solution relations become fully understood and effectively worked into the entire context to which they belong. This may again call for new acts of insight (Usher, p. 526, quoted by Ruttan, 1959, pp. 601, 602). Following Usher s theory of invention, which mainly draws on Gestalt theory, some authors have proposed new developments to a theory of invention based on another field of inquiry, the philosophy of science. Vincenti (1980) and Constant (1974) along with Dosi (1982) suggested using the concept of TPs, more specifically preparadigmatic states of technology evolution, to understand invention. This perspective adds important insights for our understanding of invention. Radical invention is about solving problems, more specifically ISP and the processes through which they emerge are non linear, e.g., they require a new technological paradigm (TP). The TP moves beyond abstract conceptualization of problem solving in the sense that it argues that innovation is about solving different types of problems. While in previous conceptualizations of technical change, innovation was viewed as problem solving, the problems remain an implicit element in this approach. There was no link between solution and problem. In Dosi (1982), the TP contains solutions that relate to problems, and problems that are explicitly stated. In pre-paradigmatic states of technology evolution, as the scientific and technological principles are not known, institutions face extraordinary problems or ill-structured problems which can lead to radical product innovation. The problems are extraordinary because solutions cannot be found within the boundaries of a TP. In paradigmatic states of technology evolution, firms face

15 normal problems which lead to incremental innovation. Solutions can then be identified within the boundaries of a TP. While these new developments emphasized non linearity methods for radical invention to occur another stream of the literature favors a more linear view. One theory of invention was proposed by Flemming (2001b) where it defines invention as the recombination of existing elements into new ones an invention can be defined as either a new combination of components or a new relationship between previously combined components (Fleming, 2001). According to Henderson and Clark (1990), even a simple rearrangement of components that are already in use, can, according to be a main cause of destabilization in key industries. To sum up, a key issue in this debate is whether radical invention comes about through linear or non linear methods. This is an unsolved puzzle that needs to be handled because it can determine the definition of alternative strategies of knowledge production than can (or not) lead to a solution to a problem. 3. On the structure of ISP : some hypotheses In order to identify the patterns of knowledge dynamics underlying the emergence of radical innovation we argue that this case of product innovation can be associated with ISP. This argument comes, on one hand, from Simon and Newell s pragmatic view on discovery (1971). Accordingly, we argue that discovery is about solving specific types of problems characterized by different levels of complexity and that to solve them we need to relate processes with outcomes (while in cognitive sciences reasoning and problem solving are regarded as distinctive processes with some degree of overlap, in the pragmatism tradition, reasoning and problem solving are combined to form one

16 single approach). On the other hand, we draw on Dosi s pioneering work on technological paradigm and, based on that view, we argue that radical invention can be associated with ISP. By combining these two views, we can gain further insights on the issue at stake, e.g., whether radical invention emerges through recombination of existing knowledge (linear view) or it requires a major mental leap (non linear view). However, we face a major challenge to pursue further this approach. The application of problem solving theory to innovation studies has analyzed the causes for large differences in difficulty to solve distinct types of problems. Based on this argument, the innovation studies literature has dealt with a variety of problems, problems about which we know ex ante its components and its interdependencies but we do not know the values attached to the variables of the problem, problems characterized by parametric uncertainty which require both information gathering capabilities or problem solving capabilities. This framework has been extended to encompass problems about which we know the components and the complexity of the problem varies with the number of components and the interdependencies within existing sub-systems, which are not know ex ante, complex problems characterized by so called recombinant uncertainty which require integrative capabilities. What remain much less clear concerns the patterns of problem solving associated with ISP. Taking into account this gap, we will combine literature on cognitive sciences with innovation studies literature on ISP to further pursue this approach. The different positions regarding the nature and ways to deal with ISP will be critically reviewed. Based on this review, some new hypothesis will be put forward. Beforehand we summarize the positions at stake with regard to ISP. Three major positions have emerged. A former position introduced by Newell and Simon (1969),

17 Simon (1969, 1973a, b) considers that, well-structured problems (WSP) and ISP define a continuum and, that ISP are a sub-class of near-decomposable problems. Similarly to complex problems, ISP are problems about which we know ex ante its components and the complexity of the problem varies with the number of components and the interdependencies, which are not know ex ante. As in ISP the interdependencies between sub-systems are weak although not necessarily negligible - it is still possible to apply hierarchic problem solving to reduce the complexity of the problem and to solve it, e.g., it is possible to apply linear problem solving methods to ISP. Another line of thought (Dosi and Egidi, 1991; Dosi et al. 1997; Levinthal, 1997, Pádua 2008, just to name a few) argues that ISP and complex problem do not define a continuum because ISP involves more fundamental lack of knowledge of the problem, the fact that our representation of the problem is not a valid abstraction of the problem (which goes against a linear view of problem solving). A third line argues that ISP involves some form of local search (Flemming, 2001, Schoenmakers W., Duysters, 2010). Although it is not clear while reading these studies about whether other properties are involved it is stated that at least local search is involved when dealing with ISP. In other words some form on linearity is involved in solving ISP On the structure of ISP Whereas in innovation studies, most attention has been given to well-structured problems, more specifically to the structure of complex problems, much less attention has been given to the structure of ill-structured problems, with the exception of the work of Cyert and March (1963) which have coined this expression, the work of Simon in his paper The

18 structure of ill structured problems (1973) and Marengo (1992). This issue will be considered at this point. Beforehand we need to provide a definition and then discussion. Well structured versus ill structured problems Simon states that an ill-structured problem is a problem whose structure lacks definition in some respect. 4 A problem may be ill-structured with regard to many facets: the goal state, the states and the initial state, the concepts or search paths 5. For example, one could have a well-defined initial state and an ill-defined goal state and set of operators ( make something pretty with these materials and tools ) or an ill-defined initial state and a well-defined final state (e.g. prove a particular mathematical conjecture). 6 But being well-defined depends on the familiarity of the problem space elements, and this in turn depends on an interaction between the problem and the problem solver. More specifically, it rests on the process of recognition. Before any search process can be applied, its relevance must be recognized by the detection of appropriate patterns in the situation. Observation of such patterns evokes information about the situation that can help guide the search. As this information usually is domain-specific, the recognition mechanism tends to make it available just where it is potentially relevant (1969). According to Simon, ill-structured problems should be defined as a residual concept (i.e. by the attributes they lack in comparison with well-structured problems). According to Simon (1973) a well-structured problem, or WSP, is defined by a list of requirements, though the 4 For Simon well structured and ill structured define a continuum of problem types and not a dichotomy. 5 In ill-structured problems, the number of goals, which are vaguely defined, must be considered in the problem-solving process (Chi & Glaser, 1985). It is uncertain which concepts, rules and principles are necessary for the solution (Wood, 1983). Ill-structured problems may possess multiple solutions, solution paths or no solution at all (Kitchener, 1983); may possess multiple criteria for evaluation of solutions, because there is no universal agreement on the appropriate solutions (Voss, 1988) and require solvers to express personal opinions or beliefs about the problem in the process of interpretation (Macham & Emont, 1989; Hong 1998); most research on ill-structured problems has focused on think-aloud protocols. 6

19 criteria are not absolute but generally express a relation between characteristics of a problem domain, on the one hand, and the characteristics and power of an implicit or explicit problem-solving mechanism, on the other hand. Well-structured problems are defined by the following elements: - there is a definite criterion for testing any proposed solution and a mechanizable process for applying the criterion; - there is at least one problem space in which can be represented the initial problem state, the goal state and all other states that may be reached or considered in the course of attempting a solution of the problem; - attainable state changes (legal moves) can be represented in a problem space as transitions from given states to the states directly attainable from them, but considerable moves whether legal or not can also be represented - that is all transitions from one considerable state to another; - if the actual problem involves acting upon the external world, then the definition of state changes and of the effects upon the state of applying any operator reflect with complete accuracy in one or more problem spaces the laws that govern the external world; -any knowledge that the problem solver can acquire about the problem can be represented in one or more problem spaces. Having defined the principal characteristics of an ill-structured problem, we need to take the next step and distinguish well-structured and ill-structured problems according to a further element: the structure of the search space. Central to Simon s perspective is the use of science of complexity to our understanding of the structure of ISP. Firstly, we provide the definition of complex systems provided by Simon (1969):

20 Complex systems are composed of a large number of parts, which interact in a non-simple way. In such systems, the whole is more than the sum of the parts, not in an ultimate, metaphysical sense, but in the important pragmatic sense that, given the properties of the parts and the laws of their interaction, it is not a trivial matter to infer the properties of the whole (Simon, 1969, p. 468). Similarly to complex problems, ill-structured problems are somewhat complex in the sense that they are characterized by a very high number of sub systems and interactions between them. The key argument of H. Simon is that complex systems evolve towards a hierarchy: By a hierarchic system or hierarchy, I mean a system that is composed of interrelated sub-systems, each of the latter being in turn hierarchic in structure until we reach some lowest level of elementary subsystem. It refers to all complex systems analyzable into successive sets of sub-systems and I speak of formal hierarchy when I refer to the more specialized concept (Simon, 1969, 468).b The central hypothesis introduced by Simon on ISP is that WSP and ISP define a continuum and, that ISP is a sub-class of near-decomposable problems. Similarly to complex problems, ISP are problems about which we know ex ante its components and the complexity of the problem varies with the number of components and the interdependencies within existing sub-systems, which are not know ex ante. It has been hypothesized that the additional source of uncertainty which distinguishes ISP from complex problems refers to the fact that the interdependencies between the sub-systems of a problem are not known ex ante. But, as in ISP the interdependencies between subsystems are weak although not necessarily negligible - it is still possible to apply hierarchic problem solving to reduce the complexity of the problem and to solve it. This perspective has been severely criticized. For example, this approach has been used to portray a type of discovery, e.g., scientific discovery. Alongside this perspective,

21 different computers programs have been designed to make discoveries (for example Kepler s law). While these models were successful in rediscovering many scientific laws, it presented important limitations, because the input data introduced in the programs were so structured that the discoveries were fairly obvious (it took four hours for a computer program to rediscover a law that took centuries to be discovered by non artificial agents). The criticism was then that then to represent a problem was to (almost) solve it. As such, although the insight presented by H. Simon that ISP have the hierarchy and decomposability was interesting, it ignored real complexity, e.g., the fact that problems present so many interdependences that it is impossible to solve them without finding the adequate knowledge representation. This criticism leads u straight away to the second position on how to deal with ISP. ii) Radical invention, ISP and non linearity The second hypothesis, introduced by Dosi and Egidi (1991), Dosi et al (1995), claims that, when we are not able to fully represent the problem, e.g., either its components or interdependencies necessary to solve a problem, as is the case for ISP, it is neither possible to decompose and solve the problem according to a given hierarchy nor possible to have access to a corresponding (given) sub-knowledge representation adequate for each sub-problem. Instead we may need additional information about the structure of the problem itself before being able to apply hierarchic problem solving otherwise we may not be able to decompose it, in the way most adequate to solve it, due to lack of critical information about its overall structure. Marengo et al (2000) point out: So far we have supposed that the structure of the problem, the representation of the problem to be searched, is exogenously given and cannot be manipulated. In this situation, we suppose that agents do not know the objective problem but only its representation and therefore that is

22 the space defined by such representation which is being searched by a decomposition scheme In order to test these propositions about ISP, NK models were used by Frenken et al(1999). Based on simulation models, their results show that: according to the degree of correlation between search landscapes, the decomposition strategy can be maximal or minimal. If there are many interdependencies among components, the maximum degree of decomposition does not generally allow the system to reach the optimal solution but only a local optimum which is close to the initial position. At the other extreme, the problem could not be decomposed at all, that is the organization could apply search algorithms which explore at the same time all N dimensions of the problem. In between, there are many possible decompositions of the problem. (Frenken et al, 1999 p.3). From this perspective, the causes for large differences in difficulty to solve distinct type of problems are twofold (WSP and ISP). A WSP results from a problem solving gap. A problem solving gap which originates from the lack of understanding of the decomposition strategy more adequate to solve a problem. A second cause originates from a more fundamental lack of knowledge of the problem, the fact that our representation of the problem is not a valid abstraction of the problem (this applies to ISP). With regard to ISP, they are hardly decomposable (e.g. there is no linearity in the process). This has been possible one of the strongest criticisms to the early models of discovery. However, although interesting it does tell how change of representations occur. Moreover it remains somehow abstract because it relies on simulation models as opposed to empirical methods. iii) Radical invention as an emergent process A third line argues that ISP involves some form of local search (Flemming, 2001, Schoenmakers W., Duysters, 2010). Although it is not clear while reading these studies about whether other properties are involved it is stated that at least local search is

23 involved when dealing with ISP. In other words some form on linearity is involved in solving ISP. This type of finding goes well (to some extent) with NK models. NK models simulate the evolution of complex systems, in which the elements function interdependently, in order to characterize the evolution of social systems. In these models, order emerges mainly through local interactions. This finding goes well along results obtained on invention (Flemming, 2001, Schoenmakers W., Duysters, 2010). Taking into account that they are supported by complexity models as well, these findings seem even more persuasive. However, and paradoxically, some questions emerge. NK models argue that there are several regimes. There are three stages of evolution, ordered regime (search is local); rugged landscape (near the edge of chaos) where search is local and concentrated around some basins of attractions and chaotic regimes. According to Kauffman (1993) rugged landscapes near the edge of chaos are more prone to change or mutation. In other words, for mutation to occur, order and instability have to co-exist. Bearing in mind, on one hand, the finding of previous studies on invention which emphasize the importance of local search (e.g., a more ordered regime) and, on the other hand, the existence of this paradox of order and instability for mutations to occur, we wonder which regime will be more prone to discovery? Summing up, while these perspectives are all interesting they lead us to mixed findings about the nature of ISP and ways to deal with them. Moreover, although these models use a variety of methods (ranging from computer simulation to empirical methods) we lack a methodology whereby we can test whether and how these properties apply to ISP

24 3.2. ISP and hypotheses The previous section argued that changes knowledge representation might be critical to solve ISPs. Accordingly, it will be argued that changes to knowledge representation will be essential to solve an ISP. However, due to the fact that there is not a process theory about how this change of representation occurs, we will draw on different elements to provide an analytical framework to describe its role in the dynamics of conversion of an ISP into a WSP. We draw on Newell and Simon s (1971) conception of problem solving as heuristic search through a problem space, combined with Simon s work on complexity (1969) and Kauffman s NK models to deal with changes of knowledge representations. On one hand, and, alongside Simon s work we will put forward the hypothesis that the analysis of complex problems goes a long way towards supporting the understanding ISP, it defines important properties which characterize knowledge dynamics in scientific discovery. On the other hand, we will suggest, alongside propositions of Dosi et al (1997) and Kauffman, other types of proprieties of knowledge production for ISP. Simon puts forward the hypothesis that ill-structured problems are a sub-class of neardecomposable problems (1973). As it is proposed that near decomposable problems evolve towards a hierarchy 7, it is proposed that ISP will also have the hierarchy property. Similarly to near-decomposable problems which are decomposable, ISP also has this property. Moreover, as near-decomposable systems are characterized by strong interactions within the sub-systems, ISP is also characterized by strong 7 It is not a hierarchy with a formal character since it is not based on authority or subordination, but the concept of hierarchy refers here to systems. By a hierarchic system or hierarchy, I mean a system that is composed of interrelated sub-systems, each of the latter being in turn hierarchic in structure until we reach some lowest level of elementary subsystem. It refers to all complex systems analyzable into successive sets of sub-systems and I speak of formal hierarchy when I refer to the more specialized concept (Simon, 1969, p196).

25 interdependencies within sub-systems. Additionally, as noted by Simon (1969), ISP can be characterized by weak but non negligible interdependencies between the subsystems. As could be expected, systems characterized by weak but non negligible interdependencies have a lower degree of decomposability (but they are still decomposable). So the system will take more time to evolve towards a hierarchy. Following Simon s position that ISP are a sub-class of near decomposable problems the following hypothesis are put forward: Hypothesis nº 1: ISP has the hierarchy property and are decomposable, Hypothesis nº 2: ISP is characterized weak but non negligible interdependencies between sub-systems Moreover, it has been suggested that when we do not know the structure of the problem ex ante, we don t have a given knowledge representation of the problem, the knowledge representation itself has to be envisaged as an endogenous factor. The production of knowledge requires the search or new state spaces which are added to the initial representation. The production of knowledge opens up new potential solutions or technological paths by pointing out new components or additional interdependencies between the sub-systems necessary to solve the problem thereby changing the structure of the search landscape itself, e.g., the guiding knowledge representation. We add the hypothesis that: Hypothesis nº 4: ISP is characterized by a non linearity : Hypothesis nº5: ISP are characterized by rugged landscape near the edge of chaos

26 The extent to which those proprieties (such as hierarchy, decomposability, and non linearity) characterize the structure and evolution of the knowledge representation will be explored empirically through one case study Data collection The methodology used in this case is inductive and relies on the mapping of dynamics of knowledge production in the HIV vaccine over a period of 17 years ( ). A semantic network characterizing the relationships between concepts in HIV vaccine search has been produced. A bibliometric analysis, more specifically co-word analysis (Callon, 1991), based on manual indexation of more than 600 abstracts, has been conducted alongside the template defined by the semantic network (Burgun A, Bodenreider O., 2001). More than 27,000 co-occurrences were obtained. The Alamos database on candidates for HIV vaccines has been used. Before proceeding to the section of result we will provide further elements on the semantic network designed for the purpose of this study. It is first necessary to consider that semantic nets are composed of two types of primitives, nodes and links (Brachman, 1977). The semantic types are the nodes in the Network and the relationships between them are the links. To build a semantic network in the HIV vaccine, one has to first consider what to put in the two types of primitives which constitute the semantic network (nodes and arcs). Taking into account that the search of the HIV vaccine research includes the interactions between the HIV virus and immunology studies, nodes represent concepts associated with the host immune responses and the HIV virus, whereas arcs represent relationships between those concepts, interactions between host immune and the virus or between viruses.

27 The primary link between the semantic types is the IsA link. The IsA link establishes the hierarchy of types within the network. For example, the semantic type research on host immune response has two children (cells and molecules) and each one in turn has several children and grandchildren. There is also a set of non-hierarchical links, grouped into 5 major categories: physically related to ; spatially related to ; temporally related to ; functionally related to ; and conceptually related to (Burgun &, Bodenreider, 2001). For example the concept genes relate to the concept proteins through a non-hierarchical link such functionally related to. The semantic designed for the purpose of this study has seven major layers. The aim of vaccine research is to find a way to intervene in the host s immune response in order to provide protection from specific infections. As vaccines are based on the introduction of components of the pathogenic organism into the host organism, research on the HIV vaccine focuses on research into the host s immune response and research into the HIV virus, as well as their interactions. In other words, search in the area of the HIV vaccine is based on searching for the HIV components which will produce the best results in terms of immunological response. Those two elements define the first layer of the hierarchy, and thus the first two branches of the semantic network. Research into the host s immune response is subdivided into two classes: innate or nonadaptive and adaptive responses. Whereas innate immunity is general, acquired immunity is specific to the foreign elements it faces and it requires previous exposure to the harmful substance which enters the body. Research into the host immune response involves research into the specific and non-specific defensive responses. The two types of immune response mentioned, research on non-specific defense and specific defense, form the second level of the hierarchy.

28 Research on the specific and non-specific immune response originates in four distinct research paths, which may lead to HIV vaccine discovery. Accordingly, there are four research programmes upon which scientific research can draw for the discovery of the HIV vaccine. Non-specific defense immunity leads to the Innate research path, whereas specific defense immune response generates three research paths called Humoral immunity, Cellular immunity, and the combination of the two Humoral and cellular immunity. Those research paths form the third level of the hierarchy. The four strategies have been designated: i) the strategy of antibody development; ii) the strategy of antibody and cell-based development; iii) the strategy of cell development; iv) the strategy based on innate immunity. The next layers refer to the experimental studies described in the papers on HIV. The experimental studies are in turn subdivided into generic immunological reactions, neutralization studies and vaccine studies. In each of these areas, researchers focus on different HIV targets (different HIV components) and assess the immune response of the host organism to those different viral components, with different criteria in mind. Firstly, researchers conduct generic immunological studies to identify the proteins (at the level of the sequence of amino-acid) with higher antigenic properties. Then they study the chemical and biological functions of those viral proteins. This defines the fourth level of the hierarchy. When the properties of the host s immune response and the pathogenic elements are identified, it is possible to move to a further level and define good targets (the type of viral protein, gene or enzyme to be used) as well as adequate immunological reactions (the type of antibodies) for neutralization studies, whether in vitro, in vivo or ex vivo. Only when knowledge on the generic properties of the host organism and the

29 pathogenic agent is accumulated can neutralization studies be conducted. In other words, the properties which are identified in the previous level of the hierarchy are inherited by the following level. This defines the fifth level of the hierarchy. When neutralization studies produce good results, both in vivo and in vitro, research is then possible to move to the next layer, vaccine technology studies. In other words, the properties which are identified in the neutralization studies are inherited by such vaccine technology studies. Vaccine trials are conducted either on animals (mice and monkey models are the most common) or humans, such as trials carried out in seropositive, seronegative at risk, and then seronegative individuals. This defines the seventh level of the hierarchy. Vaccine technology studies in turn comprise several sub-elements. One of these is the technological trajectories (which are further subdivided into 8 types). A second element is the antigens used in vaccine design. A third element is the adjuvant used for vaccine design. A fourth element is the vectors used for vaccine design. All these four elements are situated at the same sub-hierarchical level (they are sub-branches of the seventh hierarchical level). But the nature of the relationship between these sub-elements, the technological trajectories, adjuvant, antigens and vectors are non-hierarchical. It is a link of the type is associated with. The combinations between these elements define the combinatorial space upon which new vaccine trials draw. These combinations define the vaccine technological system. This semantic network served as a template to map the co-occurrences in vaccine abstracts. It is argued that by examining the dynamics of knowledge production in the case of the HIV vaccine we can say something about how much of (linear) problem

30 solving can support the solving of ill-structured problem and provide a tentative reflection on the boundaries between ISP and WSP. The first part of this question formulates the research questions whereas the second data describes the data collection strategy. 4. RESULTS 4.1. On the structure of the HIV vaccine search landscape As a first step, a more general view of the HIV vaccine search is provided. Cooccurrences were counted and 4 major periods of search appear % 80.00% 60.00% 40.00% 20.00% 0.00% % AB CTL AB+CTL INNATE Fig 1: Counting of co-occurrences by search landscape and years In order to account for this first set of empirical results it is necessary to provide a brief historical overview of research into the HIV vaccine. Research on the development of an HIV vaccine started in In the mid 1980s, it was thought that an envelope-based vaccine would be quickly made using recombinant DNA technology. 8 This was based on the finding that the outer envelope of HIV was important for stimulating neutralizing antibodies. 8 An envelope-based vaccine that would trigger an immune response via neutralizing antibodies using recombinant DNA technology (McMichael, 2003).

31 The graph shows that in the first period scientists chose the antibody approach to design a vaccine. Most conventional approaches of vaccine design are based on the induction of antibodies that attach to the virus and neutralize it. This is because the antibody is the only component of the adaptive immune response that can neutralize a virus particle prior to infection of a cell and it is the only immune response associated with protection for any currently licensed vaccines. Previous and current vaccines have proved successful for combating diseases such as poliomyelitis, measles and influenza. However this approach led to poor solutions for the discovery of an HIV vaccine. At a certain point in time scientists changed their interest and started giving more importance to another alternative, the CTL approach. Unlike antibodies, cytotoxic T cells recognize infected cells rather than the infectious agent itself. T lymphocytes detect the presence of a foreign substance by way of surface proteins called T-cell receptors. The CD8 T lymphocytes, also called T killer lymphocytes, eliminate cells that display antigens on their surface. This is accomplished by releasing cytotoxic substances that will rupture the cell s cytoplasm leading to its destruction (through granzyme and perforin). Very few vaccines have been made based on this approach. They are mostly experimental or less effective (like vaccine BCG against tuberculosis). This approach also led to poor solutions for the discovery of an HIV vaccine. Different vaccine designs were used based on DNA and other approaches to vaccine design. At a later stage another starting point considered was to use both immune antibody and cell mediated responses. Reflecting the fact that there are two main immune responses to pathogenic agents, scientists could choose one of two major options to start tackling this problem. There are two main immune responses, the humoral and cellular immunity, and there is often debate and speculation about which component of the

32 adaptive immune system is most important for immunity. Many vaccines rely on both approaches. But the results were not very encouraging either. In other words scientists and engineers tried to put two things which do not work well together. All sorts of vaccine design were used by that time, peptide, sub-unit, DNA vaccine etc. A fourth approach was the search for radical new strategies for vaccine discovery (2002 onwards): the humoral and cellular research path versus the innate/acquired path as distinct explanatory hypotheses to account for correlates of immunity within the scientific discovery search space. Vaccine designs are based on the search for radical new scientific approaches and new combinations of technological heuristics due to persistent functional failure. Again all sorts of vaccine design were experimented. The second finding is that results group around 4 major clusters. Findings show that four major clusters emerge over time. Fig 2: Vaccine search and clusters Figures on cluster here The third finding is that the knowledge representation used to guide search evolves towards a hierarchy. The graphs in the following section show that, in the first period ( ) co-occurrences are concentrated in the 3 layer of the hierarchy (for the antibody search landscape). Graph 3 shows that in the period co-occurrences start to increase for layer 4 (neutralization studies) whereas for the fourth period the

33 next layer, vaccines studies, increase. The same order of progression applies to the other search landscapes. global neutralization related vaccine related ab ctl ab+ctl vaccine + neutralization Graph 3 Counting co-occurrences for the period by strategy of search and hierarchical level global neutralization related vaccine related ab CTL ab+ctl vaccine + neutralization Fig 4 Counting for the period by strategy and hierarchical level

34 global neutralization related vaccine related AB CTL ab+ctl vaccine + neutralization [NO DATA AVAILABLE] Fig 5. Counting of co-occurrences for the period by strategy and hierarchical level global neutralization related vaccine related ab CTL AB+ctl Innate vaccine + neutralization Fig. 6 Counting for the period by strategy and hierarchical level Throughout these 4 periods it is possible to observe that interdependences between branches increase (counting s for the strategy that combines both antibody and CTL search increase). Whether this means that the HIV vaccine is a near decomposable problem is not clear. It seems that the problem keeps some structure (there are clear strategies of search). However, one can say that the search landscapes become more and