AI and Philosophy of Science

AI AND PHILOSOPHY OF SCIENCE
Sakir Kocabas*

Istanbul Technical University
Department of Space Sciences and Technology

Abstract
Recent research in the computational study of scientific discovery has revealed a number of critical aspects of science overlooked by the conventional philosophical study. A series of computational models developed by AI scientists to study the different aspects of historical discovery indicate that hypothesis formation, testing and verification are only a small part of scientific research. This paper investigates from AI perspective, scientific creativity, the processes of scientific research, the dimensions of scientific research, and the role of knowledge in scientific research.

------------
* Also affiliated with: Department of AI, Marmara Research Center, PK 21, Gebze, Turkey.

1. Introduction
Scientific discovery and creativity has, in the last fifteen years, become one of the special concerns of artificial intelligence (AI). Within this period, a number of research papers and two important books have appeared on scientific discovery (see, Langley, Simon, Bradshaw, & Zytkow, 1987; Shrager & Langley, 1990). Closely related with the subject, several other publications have appeared. These include one on the computational philosophy of science (Thagard, 1988), one on theory revision in science (Darden, 1991), and another one on creativity (Boden, 1990).

Langley et al.'s (1987) work posed the first serious challenge to the conventional study of science by proposing that, far from being mysterious and unexplainable, scientific discovery (and by implication, scientific creativity), can be explained in terms of a series of processes. Their work also described several computational models in support of the authors' view. Shrager and Langley's (1990) later study introduced new methods for the study of scientific development, and explained how the methods of the computational study of science were superior to those of conventional philosophical studies. Boden's (1990) work on the other hand, extended some of these views and discussed, from a cognitive scientist's perspective, how creativity in arts and literature, as well as in science could be studied within a computational context, in a more systematic way.

However, previous work leaves some important issues in discovery untouched, such as the elements of scientific creativity, the types of scientific discovery and creativity, and the dimensions of scientific research. In this study, we examine the basic cognitive concepts of creativity, and describe how these concepts are interrelated, and then discuss the role of background knowledge and the kinds of knowledge necessary for scientific research. Finally, we discuss the types of scientific discovery and the elements of scientific research, and conclude with a summary.

2. Intelligence and Creativity in Science
Creativity and intelligence are closely related concepts, so that the any attempt that brings clarity to one concept will be helpful to define the other. AI scientists rely on computational terms in their definitions. Lenat and Feigenbaum (1987) define intelligence in terms of "search", as the power to find a solution to a problem in an large search space. Later, Feigenbaum defined intelligence in terms of "knowledge assembly" rather than "search" (see, Engelmore & Morgan, 1988, vii). According to this definition, an intelligent system has the ability to assemble the neccessary body of knowledge to conduct a complex task.

However, these definitions do not capture the complexity of the concept of intelligence. A more detailed definition has been given by Hayes-Roth (1993) within the context of intelligent agents, where the author discusses the agent characteristics in three system components: perception, cognition and action. Accordingly, each component must be capable of operating independently in a coherent way. Additionally, each component must meet a series of criteria in order to be called intelligent. These criteria require that an intelligent system be capable of perceiving, thinking and acting in real-time, asynchronously, selectively, coherently, flexibly, responsively, robustly and timely. It must also be capable of developing its abilities by adaptation and learning. Many AI scientists discuss learning at two levels as symbolic and subsymbolic, and classify symbolic learning into several different types as rote learning, learning by instruction, inductive learning, deductive learning, and learning by analogy.

Creativity can also be classified into different types. Accordingly, a distinction can be made between scientific creativity and other types of creativity such as artistic, architectural, musical and literary creativity. The former may involve the discovery of a new substance, the invention of a new mechanism or method, or the construction of a new model of reality (a hypothesis or a theory). The latter however, mostly manifests itself as a work of art or a new style, and the term "creativity" is usually associated with this type.

Scientific creativity can be distinguished from other forms of creativity such as in arts, music and literature, by its extensive reliance on background knowledge and experience in history. This may explain why we do not see child prodigies in creative science as we see in music and arts. Therefore, when we talk about scientific creativity, it is to be understood within this perspective.

Scientific creativity can be investigated through five basic cognitive and computational concepts:

1) Motivation for scientific research.
2) Ability to correctly formulate research problems within a body of knowledge.
3) Ability to create a comprehensive search space for the solution of a scientific problem.
4) Ability to assemble (or induce) and implement a set of heuristics to reduce the search space.
5) Patience and stamina for the exhaustive search for solving the selected scientific problem within the constrained search space. Fig. 1 summarizes the links between these concepts. Any missing link between them, can hinder scientific creativity.


Motivation for Formulate Generate Reduce Conduct
Scientific --> Research --> Search --> Search --> Exhaustive
Research Problems Space Space Search

Fig. 1. Problem formulation and search in scientific discovery.

As indicated in the above list, research motivation tops the requirements for scientific creativity. Motivation itself can be dependent on basic psychological needs. Various types of human motivation have been studied by psychologists in the last five decades (see, e.g., Maslow, 1966). Metaphysical commitments and ontological assumptions about the world may also affect motivation (see, e.g. Kuhn, 1970, p.41). This is an important issue, but is outside the scope of this study.
Problem formulation is the second major issue in scientific research. In modern scientific research, an access to a large and systematic body of knowledge is necessary for correctly formulating scientific problems. The accurate formulation of research problems requires a mastery of the conceptual structure of the field of science involved. The creative scientist can change this structure for reformulating a research problem in his/her search for a solution. In some cases, changes in the conceptual structure involve the most fundamental concepts and principles, such as time and measurability in physics. Changing representations on the other hand, provides alternative views of the problem space, and is considered as one of the most influential parameters of creativity in science (see, e.g., Simon, 1992; Karmiloff-Smith, 1990).

Extensive knowledge may also be used in creating a comprehensive search space for the selected research problem. The search space is then reduced to a manageable size, by selecting and applying appropriate search strategies, methods and heuristics. This is necessary to reach for a solution within acceptable limits of time and resources. Once the problem is defined and constrained, exhaustive search needs to be carried out within the search space, until a conclusion is reached about the solution of the scientific problem. Scientific creativity exhibits itself during the completion of a series of research tasks. Different types of knowledge may be used for each task, as will be explained next.

3. Types of Knowledge Used in Research, Types of Scientific Discovery
Modern scientific research is one of the most complex human activities, requiring the use of different types of general and specific knowledge. The knowledge necessary for modern scientific research can be divided into four types as a) Commonsense Knowledge, b) Technical Knowledge, c) Theoretical Knowledge, and d) Methodological Knowledge (*).

Commonsense knowledge is simple, general and relatively unstructured knowledge about the world. Technical knowledge can be defined as the knowledge about instruments, methods and processes. Theoretical background is helpful, but not always essential, in acquiring this kind of knowledge. Technical knowledge can be descriptive as well as prescriptive.

-----------------------
* Knowledge used in the proecesses of scientific discovery is by no means limited to the four types listed here. There can be other types of knowledge, including religious symbolisms, to play a role in scientific research, as can be seen in the recent history of quark theory in particle physics.

Theoretical knowledge is structured, descriptive knowledge about the world, embodying classifications and numerous interrelated hypotheses. Typical examples of theoretical knowledge are the classical mechanics and electro-magnetism.

Methodological knowledge, on the other hand, is exclusively prescriptive; it can be represented as condition-action rules. Methodological knowledge includes knowledge about how to distinguish between scientifically interesting and uninteresting phenomena, how to choose between alternative goals, strategies and methods in scientific research, how to design experiments, how to propose new hypotheses, and how to generalize, test and evaluate them. It is mostly the extent of this type of knowledge that makes the difference between a research scientist and a nonscientist.

Unlike the inference rules in theoretical knowledge, many of the methodological rules rely on extralogical methods such as inductive generalizations, abduction, abstraction and analogy. Such rules are frequently used in formulating problem states, in constraining large search spaces, and in hypothesis formation during the activity of scientific research.

Scientific creativity can be examined in relation to the scope of the research in which a discovery takes place. Kocabas (1992c) introduces a classification of scientific discovery as follows: 1) Logico-Mathematical Discovery, 2) Formal Discovery, 3) Theoretical Discovery, and 4) Empirical Discovery. This classification is based on the categorization of descriptive knowledge by Kocabas (1992a), and reflects the types of knowledge used in scientific research, and the type of knowledge discovered. All these four types of discovery have been studied in AI by a series of computational models.

According to this classification, logico-mathematical discovery takes place, as the name suggests, in the abstract domain of logic and mathematics. The distinguishing characteristic of logico-mathematical discovery is that, in principle, it does not require experimentation or observation. Nor does it need the knowledge of a physical domain par se, except for analogical transference in some cases.

Formal discovery takes place in a formal domain involving abstract entities, their classes and properties. Formal discovery requires logico-mathematical knowledge as background knowledge, for deductive inference on formal knowledge.

Theoretical discovery requires logico-mathematical, formal and theoretical knowledge, and in general, results from theoretical analysis and synthesis. In the history of science there are rather important theoretical discoveries or inventions such as Maxwell's equations and the Einstein-Lorenz transformations.

Empirical discovery requires experimental and observational data, as well as logico-mathematical and formal knowledge. Theoretical knowledge has not been a prerequisite in the early empirical discoveries in the history of science (e.g. in the 17th and 18th century chemistry), but in modern empirical research such as in oxide superconductivity and fusion experiments, extensive theoretical domain knowledge is necessary.

4. Computational Models of Discovery
In parallel with the types of discovery described above, computational models developed by AI scientists can be classified in the same types as Logico-mathematical Models, Formal Models, Empirical Models, and Theoretical Models.

Some of the earliest AI systems such as Logic Theorist were logico-mathematical discovery models designed to prove theorems in logic. Among the more recent computational models, AM (Lenat, 1979) constitutes an outstanding example for mathematical discovery.

Lenat's (1983) EURISKO, in its applications to Naval Fleet Design, Evolution, and 3-D circuit design, can be cited as a typical example to formal discovery systems.

Some computational models of theoretical discovery are PI (Thagard & Holyoak, 1985), ECHO (Thagard & Nowak, 1990), GALILEO (Zytkow, 1990), and PAULI (Valdes-Perez, 1994). The first two could better be characterized as concept discovery systems, and as such, are closer to formal discovery models. GALILEO on the other hand, is an interesting example of discovery by theoretical analysis in that it discovers more expressive forms of scientific laws. The PAULI system is another interesting model which has led to the discovery of a general theorem about the quantum values of elementary particles in physics.

Empirical discovery is an extensively studied area in AI, and a number of computational models have been designed to investigate its various aspects. Empirical discovery systems can be divided into two main classes as qualitative and quantitative models, although this distinction is sometimes irrelevant. Among the qualitative discovery systems, GLAUBER (Langley, et al., 1987) models the discovery of the acid-base theory in the 17th century chemistry. STAHL (Zytkow & Simon, 1986) and STAHLp (Rose & Langley, 1986) simulate the discovery of the componential models in the 18th century chemistry, the latter with the additional capability of partially modeling the paradigm shift from the phlogiston theory to the oxygen theory. AbE (O'Rorke, Morris & Schulenburg, 1990) provides a more detailed simulation of the transition from the phlogiston theory to the oxygen theory, demonstrating the role of abductive inference in the process. KEKADA (Kulkarni & Simon, 1988) simulates the discovery of the urea cycle in biochemistry by Krebs in the 1930s, by treating the process as search in several search spaces. COAST (Rajamoney, 1990) on the other hand, treats physical systems as "scenarios", and considers theory revision as incremental changes in qualitative schemas (Forbus, 1984).

Some of the other systems are BR-3 (Kocabas, 1991) and BR-4 (Kocabas & Langley, 1995) which model the discovery of several conservation laws about the elementary particles, the latter with the ability to simulate the discovery of the neutrino in particle physics. When faced with inconsistent solution states or new evidence, both systems can revise their domain theories incrementally. PAULI (Valdes-Perez, 1994) considers certain discovery problems as matrix operations in two search spaces, and reproduces BR-3's results, together with a set of alternatives, and additionally leads to a general theorem in particle physics. MECHEM (Valdes-Perez, 1995) discovers new pathways for a set of cathalytic chemical reactions, alternative to the ones known by chemists today.

Among the quantitative discovery models BACON (Langley, et al., 1987), FAHRENHEIT (Zytkow, 1987) and IDS (Nordhausen & Langley, 1987) can be cited as prominent examples. BACON was the first successful model of quantitative discovery, which also has attracted the interest of philosophers of science(*). The IDS system on the other hand, integrates qualitative and quantitative methods.

5. Aspects of Scientific Research
Research in the computational study of science indicates that the conventional philosophical study has in its history overlooked a number of critical aspects of science. Basic differences between the computational and the conventional philosophical approaches have been described by Shrager and Langley (1990). According to these authors, the conventional philosophical tradition focuses on the structure of scientific knowledge and emphasizes the evaluation of laws and theories, while the computational approach focuses on the processes of scientific thought, and emphasizes scientific discovery including the activities of data evaluation, theory formation and experimentation.

The distinction can be extended even further. Computational study of science concerns not only with the issues of hypothesis formation, testing and verification, but also a series of other related issues. Kocabas (1992b) names more than a dozen different major tasks involved in scientific research. These are: Formulating research goals, selecting research goals, defining research framework, gathering knowledge, organising knowledge, selecting research strategies, methods, tools and techniques, proposing experiments, designing experiments, selecting experiment materials, setting expectations, conducting experiments, data collection, data evaluation, hypothesis formation, theory formation, theory revision, goal satisfaction control, and producing explanations.

Each of these research tasks may involve activities dealing with a variety of planning, classification and evaluation problems. Kocabas (1992b) provides examples from the research in oxide superconductivity for the diversity of the activities involved in these research tasks. Consider, for example the formulation of scientific research goals, choosing between formulated goals, proposing strategies, proposing experiments, and hypothesis formation.

Heuristics about formulating research goals have been studied by Kulkarni and Simon (1988), Lenat (1983), and Darden (1987). Kocabas (1992b) divides research goals into two general forms that may overlap: Those that aim at explaining a phenomenon, and those that aim to study a penomenon. Creative scientists seem to utilize several general rules for formulating their research goals. One such rule is to focus attention to problems and phenomena that have not been explained or are unexplainable within the current scientific framework. However, such problems must have some general and important implications to be worthy of investigation.

--------------------------------------------
* See, e.g. the special issue (Vol 19, No 4) of Social Studies of Science.

Some scientific research problems may be strongly related to important technological needs. Energy conversion, storage, and transfer are still major technological problems that motivate scientific research into such areas as "cold fusion", oxide superconductivity, and electrochemistry. However, interestingness in itself is not a sufficient criterion for a phenomenon to attract the attention of the creative scientist. The research goals that are formulated must also be achievable.

It is not unusual that a scientist formulates alternative research goals in relation to a certain phenomenon. In such cases, the selection of a research goal among alternatives is another research task. Scientists use several selection criteria in deciding which problem to primarily focus on. Some of these constraints conflict with one another, and resolving such conflicts may not be a trivial task for the scientist.

Selecting research strategies is another important task for accomplishing a research goal. Strategy selection depends on the type of the research goal, such as explaining or examining a phenomenon. If the research strategy inmvolves experimentation, then the type of experiments needs to be decided.

Once the experimentation strategy is selected, the scientist has to decide about the relevant processes and techniques for the current strategy. S/he also has to decide about the experiment materials, and has to classify these materials against a set of parameters such as availability, likeliness to yield success, cost and relative hazards (e.g., radioactivity, flammability and corrosiveness), and select the best materials for the experiments.

Scientific experiments need to be designed and conducted according to certain theoretical frameworks, observation and measurement standards and procedures. Experimental variables must be defined beforehand, for tests are carried out to measure the variations between the values of these variables. The experimental data is evaluated to make sure if they reflect any violation of the experimental conditions. Hypotheses are formed or revised only after data evaluation.

Hypothesis formation is one of the most important tasks of scientific research. Despite the fact that it has been a primary concern of the conventional philosophy of science for a long time, it still needs a detailed investigation. In our study on oxide superconductivity research (see, Kocabas, 1992b), we have identified over 40 hypothesis formation heuristics that were utilized by scientists working in this field. The majority of these heuristics are general, while some are domain specific.

The diversity of interrelated research tasks is by itself sufficient to show that, scientific discovery is not a logical procedure or a process in itself, but the product of a series of complex processes called scientific research. Scientific creativity may be required in any of the research activities in these processes. History of modern physics has numerous examples of these processes. Although an extreme example, consider the design, construction and the operation of the CERN particle accelerator, where research involves proposing and designing experiments, setting expectations, conducting experiments, data collection, data evaluation, hypothesis formation, verification, and theory revision.

Computational models continue to be developed for modeling different aspects of scientific research. One of the hopes of research in this direction is to be able develop complete models for research, or artificial research assistants capable of directing research in different fields of science.

The increasing use of AI techniques in computational modeling may culminate in diminishing the role of mathematical reasoning in simulation. It does not seem unreasonable to expect complex physical systems (including physical theories themselves) be represented as computational models.

Computational modeling may provide other advantages in theoretical analysis and theory revision, for the use of such models can make intractably complex theories easier to grasp. Similarly, the use of computational models can make scientific explanations more systematic, more accurate and correct to the point.

6. Conclusion
Conventional philosophical approach ignores the multiplicity of the tasks and activities involved in scientific inquiry. We believe that, a much more detailed and careful examination and analysis of science is needed than that is envisaged by the conventional study of science. The computational approach provides both the necessary concepts and methods for such a study.

References
Boden, M. (1990). The creative mind. Sphere Books, London.

Darden, L. (1987). Viewing the history of science as compiled hindsight. The AI Magazine, 8, No. 2, 33-42.

Darden, L. (1991). Theory change in science: Strategies from Mendelian genetics. Oxford University Press, N.Y.

Engelmore, R. and Morgan, T. (1988). Blackboard systems. Addison Wesley.

Forbus, K.D. (1984). Qualitative process theory. Artificial Intelligence, 24, 85-168.

Hayes-Roth, B. (1993). Architectural foundations for real-time performance in intelligent systems. In David, J-M., Krivine, J-P., and Simmons, R. eds, Second Generation Expert Systems. Springer-Verlag, New York.

Karmiloff-Smith, A. (1990). Constraints of representational change: Evidence from children's drawing. Cognition, 34.

Kocabas, S. (1991). Conflict resolution as discovery in particle physics. Machine Learning, Vol 6, No 3, 277-309.

Kocabas, S. (1992a). Functional categorization of knowledge. AAAI Spring Symposium Series, 25-27 March 1992, Stanford, CA.

Kocabas, S. (1992b). Elements of scientific research: Modeling discoveries in oxide superconductivity. Proceedings of the ML92 Workshop on Machine Discovery, 63-70.

Kocabas, S. (1992c). Evaluation of discovery systems. Proceedings of the ML92 Workshop on Machine Discovery, 168-171.

Kocabas, S. and Langley, P. (1995). Integration of research tasks for modeling discoveries in particle physics. In Working Notes of 1995 Spring Symposium Series, AAAI Press, CA.

Kuhn, T.S. (1970). The structure of scientific revolutions. The University of Chicago Press, Chicago.

Kulkarni, D. and Simon, H. (1988). The processes of scientific discovery. Cognitive Science, 12, 139-175.

Langley, P., Simon, H., Bradshaw, G., and Zykow, J. (1987). Scientific discovery: Exploration of the creative processes. MIT Press.

Lenat, D.B. (1979). On automated scientific theory formation: A case study using the AM program. In Hayes, J., Michie., D., and Mikulich, D.I. eds., Machine Intelligence, 9, 251-283, Halstead, New York.

Lenat, D.B. (1983). EURISKO: A program that learns new heuristics and domain concepts. Artificial Intelligence 21, 61-98.

Lenat, D.B. and Feigenbaum, E. (1987). On the thresholds of knowledge. Proceedings of the Tenth International Joint Conference on Artificial Intelligence, 1173-1182.

Maslow, A.H. (1966). The psychology of science: A reconnessaince. Harper and Row Publishers, N.Y.

Nordhausen, B. and Langley, P. (1987). Towards an integrated discovery system. Proceedings of the Tenth International Joint Conference on Artificial Intelligence, 198-200.

O'Rorke, P., Morris, S. and Schulenburg, D. (1990). Theory formation by abstraction. In Shrager, J., and Langley P. eds. Computational models of scientific discovery and theory formation. Morgan Kaufmann, San Mateo, CA.

Rajamoney, S.A. (1990). A computational approach to theory revision. In Shrager, J., and Langley P., eds., Computational models of scientific discovery and theory formation. Morgan Kaufmann, San Mateo, CA.

Rose, D. and Langley, P. (1986). Chemical discovery as belief revision. Machine Learning, 1, 423-452.

Shrager, J., and Langley, P. Eds. (1990). Computational approaches to scientific discovery. In Shrager, J., and Langley P., eds., Computational models of scientific discovery and theory formation. Morgan Kaufmann, San Mateo, CA.

Simon, H.A. (1992). Scientific discovery as problem solving: Reply to critics. International Studies in the Philosophy of Science 6(1): 69-88.

Thagard, P. (1988). Computational philosophy of science. The MIT Press, Cambridge, MA.

Thagard, P. and Holyoak, K. (1985). Discovering the wave theory of sound: inductive inference in the context of problem solving. Proceedings of the Ninth International Joint Conference on Artificial Intelligence, 610-612.

Thagard, P. and Nowak, G. (1990). The conceptual structure of the geological revolution. In Shrager, J., and Langley P., eds., Computational models of scientific discovery and theory formation. Morgan Kaufmann, San Mateo, CA.

Valdes-Perez, R.E. (1994). Algebraic reasoning about reactions: Discovery of conserved properties in particle physics. Machine Learning 17 (1), 47-68.

Valdes-Perez, R.E. (1995). Machine discovery in chemistry: New results. Artificial Intelligence 74 (1), 191-201.

Zytkow, J.M. (1987). Combining many searches in the FAHRENHEIT discovery system. Proc. 4th Internatonal Workshop on Machine Learning, Morgan Kaufmann, CA. 281-287.

Zytkow, J.M. (1990). Deriving laws through analysis of processes and equations. In Shrager, J., and Langley P., eds., Computational models of scientific discovery and theory formation. Morgan Kaufmann, San Mateo, CA.

Zytkow, J.M. and Simon, H. (1986). A theory of historical discovery: The construction of componential models. Machine Learning, 1, 107-137.