The earliest good example of this paradigm was Lenat's AM (creat:lenat1) system. AM was a math discovery system. Using a base set of knowledge along with a pool of modification heuristics, AM was able to discover an astonishing range of mathematical results, some of which were unknown to Lenat. The model was guided by a collection of ``interestingness'' heuristics which helped to choose what avenues it was going to explore. While not created in order to be a creative behavior system, AM is regarded as an example of such. Using its knowledge, it was able to arrive at mathematical concepts which were novel to its own perspective. The use of interestingness was a significant proposal which future systems have also incorporated. The creative understanding theory embedded in the overall ISAAC approach to reading also uses interestingness as a guide to the creative process. However, the ISAAC approach to creativity is a much more directed one than AM. AM would generate numerous possibilities to explore; many of these would be filtered by Lenat himself. In other words, Lenat was an integral part of the control loop. In the ISAAC view of creativity, the process of generating concepts is very different from the AM approach. The process is also more directed and constrained than in the AM work. This leads to fewer bizarre ideas being generated and no need for a human to then discard them.
BACON ([#!creat:langley1!#]) is another discovery program which has been described as possessing creative potential. BACON was a series of data-driven models which attempted to discover scientific principles. Given a set of input variables, BACON would search for an equation set which explained the data. If a hypothesis is proposed which requires more information, BACON will propose an experiment in order to gather the additional data. With this technique, BACON was able to discover scientific knowledge such as the ideal gas law and Kepler's laws of motion. BACON's model of creativity was mainly an implicit one embodying several interest heuristics which guided its processing. Examples of these are strive for a linear data fit and manipulate constants.
Arguing that systems such as BACON were too limited in their scope and knowledge, Kulkarni and Simon developed the KEKADA system (creat:kulkarni1), which attempts to model Krebs' discovery of the urea cycle. The heuristics used by KEKADA were developed through careful analysis of the notes left by Krebs detailing the methodology he went through to arrive at his discovery. Although intended to model a particular discovery by a given scientist, KEKADA does contain general-purpose knowledge which could be applied to other discoveries, although this has not been done at this time. Finally, KEKADA's heuristics include ones such as consider alternative explanations and propose experiment.
The creative understanding portion of the ISAAC theory also relies on control heuristics to guide the process; that is the contribution of this facet of previous work--the realization of a loose reasoning process bounded by control heuristics. There is a difference in the level of these, however. In the ISAAC theory, the heuristics are not bound to a particular domain as with BACON and KEKADA; instead, the heuristics describe the types of conceptual manipulations which may be performed. In ISAAC, in addition to utilizing different methods to achieve creative behavior, the domain specific information is injected into the process by the control supertask, rather than within the creative understanding process itself. As a result, I argue that the creative understanding process is more general than the type of creativity modeled in the scientific discovery theories--their theories of creativity were too domain-specific.
While a great deal of work has been concerned with scientific discovery and creativity, there is at least one computer model which attempts to provide a theory of artistic creativity. The AARON model of Cohen (creat:cohen1) is a computer model which creates artistic productions. Cohen himself was an established artist before embarking on his computer explorations of creativity and artistic production. AARON uses a production system of artistic ``rules'' which guide its drawing process. These are used along with the given context (which includes a style of art desired) to produce artwork which is considered by most to be fairly artistic. Unfortunately, the source of power in the AARON system is not fully understandable by anyone (with the possible exception of its creator). Further, the rules it possesses execute without the system giving any critical thought to the problem of what it is producing--in other words, AARON has no awareness of what it is doing or how it is doing it, it simply does what the rules tell it to. So, while it represents an artistically creative computer model, it is unclear exactly what we can learn from such an endeavor, as AARON now stands. It is simply too atheoretical.
There are, of course, researchers who are concentrating on creativity, rather than focusing on discovery systems. These researchers attempt to discover the processes by which creativity may exist. Consider the work of Hofstadter and his research group. Several significant systems have arisen from their work, including COPYCAT, TABLETOP, and Letter Spirit (see, for example, [#!creat:hofstadter1!#,#!creat:mcgraw1!#]). The models embody two significant ideas which are claimed to lead to creativity. First, there is no central executive coordinating the creative behavior. Instead, there are numerous small processes which are chosen in a probabilistic fashion from a pool of all such processes. Second, concepts in the models exist in an organization scheme known as the SlipNet. This is a way of organizing knowledge such that semantically similar ideas can ``slip'' when needed. So, while attempting to work with a given concepts, other concepts may be slipped into the processing behavior, allowing novel results to occur. While each project has interesting differences, they all share a common notion of computational creativity as described here. And, while my research is from a different background, my work recognizes the ability for concepts to alter as required by the problem. The Hofstadter approach is more subsymbolic than my own; I believe my knowledge representation (described in Chapter 4) gives my theory the capability of this ``slippage,'' just simply in a different manner than the SlipNet method. But, unlike the Hofstadter approach, the ISAAC system makes a commitment to a much higher-level control of the creativity process. Also, the ISAAC theory claims that the four tasks of memory retrieval, analogical mapping, base-constructive analogy, and problem reformulation interact to give rise creative behavior, rather than the single technique of conceptual slippage. And, since ISAAC always has a comprehension goal guiding the creative aspects of the system, there is a principled bounds on the process generated by the system itself.
At a more symbolic level of description, Kolodner and Penberthy (creat:kolodner1) extended the standard model of case-based reasoning to make the technique more conducive to the creation of creative solutions. In this case, the domain used was unusual meal planning as performed by the Creative JULIA system. They propose that, in a problem solving situation, a reasoner retrieves a large number of potentially relevant cases. These are used in their entirety, as well as providing case pieces which can be combined to form new solutions. The theory also makes provisions for evaluation to be a primary operation, and it allows for the inclusion of potentially bizarre cases which (in straightforward case-based reasoning) may have been thrown out. Thus, the work combines an analytic approach and a generative approach in order to produce a better understanding of the problem to be solved (in this case, the meal to be designed), thereby allowing it to produce a set of possible solutions, which can then be evaluated for how well they fit the original criteria. In this work, there is no ``creativity process'' which needs to be discovered; instead, what is needed is the discovery of how everyday processes interact to produce creative behavior. My work is a logical extension of this in several ways. First, the theoretical claim that creativity is not a privileged process is crucial to the ISAAC approach. Second, the ISAAC theory of creative understanding utilizes a case-based reasoning approach with regards to knowledge storage, retrieval, and application. Finally, the idea that pieces of cases may be combined to handle novel situations is a key aspect of my research, as seen in the ability to dynamically create new concepts as needed. My work differs in the specification of the entire creative understanding process, showing exactly how memory, analogy, dynamic creation of concepts, and problem reformulation are integrated. My control mechanisms are different; as stated earlier, they are more bound to the conceptual manipulation level rather than to the domain level. Finally, my work is focused on understanding rather than on invention.
This idea of creativity and everyday reasoning being produced by the same underlying principles is one which is becoming increasingly more popular. Most researchers are now searching for the proper combination of ``mundane'' processes and sophisticated control heuristics which will produce creative behavior, in the proper circumstances (see, for example, [#!creat:ram1!#]). As a result, creativity is not something which is turned on or off in a model; instead, it is always active. Sometimes, the synergy of the various processes produces creative results; much of the time, the processes simply produce what would be called everyday behavior.
This philosophy of ``spontaneous'' creativity is embodied in the next model to be examined, another case-based approach to the creative process, proposed by Wills and Kolodner (creat:wills1,creat:wills2). Their case-study of an actual design project resulted in several insights into the creative process, which they are now attempting to model. Their model explains creativity as a combination of everyday processing elements. Through evaluation and problem reformulation, creative results are possible. Also, serendipitous recognition plays an important role in the theory. This is the recognition of a new situation as an instance of something being looked for and is somewhat an extension of case-based reasoning's idea of situation assessment. As the researchers point out, traditional case-based reasoning models do not live up to the potential of the theory; instead, they generally rely on mundane adaptations of known cases, rather than attempting more creative behavior.
The Wills and Kolodner work was an extension of the earlier described work by Kolodner and Penberthy. The line of research continued to be refined with the research of Simina and Kolodner (creat:simina-kolodner-1997) and their ALEC system. ALEC is a case-based design system intended to demonstrate creative behavior. The system is notable for its close integration of understanding into the design cycle--indeed, the understanding portion of the work is in part based on my own creative understanding theory (Chapter 6). However, there are differences between the two systems. Most importantly, ISAAC is an understanding system which recognizes that novel concepts must be generated while involved in an understanding (during text comprehension); ALEC is a design system which recognizes that novel concepts must be understood while involved in a design attempt. Also, the mechanism for generating novel concepts is different in their system than in mine.
Finally, Turner (creat:turner1) argues that the correct path to creativity is often found not in an explicit reasoning method but in a new way of looking at memory retrieval. Imaginative memory is the term coined by Turner to describe the creative mechanism within his MINSTREL system. This technique places the responsibility of the alterations to known objects in the control of the memory system. For example, consider a scenario in which a reasoner is attempting to remember a scenario in their past dealing with knights killing dragons. Their memory contains no such scenarios, but they are reminded of a time when a knight killed a troll, a time when a king killed a dragon, and a time when a knight wounded a dragon and then negotiated with it. These are all three variations of the original specification (dragon becomes troll, knight becomes king, and kill becomes wound). Turner's system makes incremental changes to the initial specification in an attempt to retrieve useful cases from memory. The idea of incremental changes is also a crucial idea in my own work--the control process for creative understanding is iterative with each cycle making incremental advances toward a solution. As soon as a workable solution is discovered, the process is halted, even though it may later be shown that this first solution is not correct. This incremental refinement is one element which allows my control heuristics to be more general than in many previous theories--although they exist at a level which allows bizarre concepts to result (novel concepts which have no usefulness), this is limited by the incremental nature of the process. MINSTREL differs from ISAAC in two important ways. First, MINSTREL is implemented with the stance that imaginative memory is the only technique which is necessary. Although it is certainly a powerful one, it is unclear if that technique alone can produce revolutionary novelty; it does not have the ability, for example, to combine multiple concepts to create new ones. Second, the MINSTREL system makes use of a relatively small set of control heuristics which Turner developed. These limit the imaginative memory by restricting the types of modifications which are permitted; in part, they depend on a somewhat careful placement of concepts into memory. The ISAAC system does not enforce a set of specific control heuristics; instead, it allows certain bounding conditions to arise through the interaction of the ontology and the general control heuristics which are in place (see Chapters 4 and 6 for a full discussion of this).