Kevin B. KorbKevin’s research is in artificial intelligence and the philosophy of science and the interrelation between the two, and especially the automation of scientific induction, or causal discovery. He is also co-founder of Psyche: An Interdisciplinary Journal of Research on Consciousness.

Recent presentations: The Philosophy of Computer Simulation, an invited talk at the 13th International Congress of Logic Methodology and Philosophy of Science, Beijing, 9-15 August, 2007 .

Two Technical Reports

Kevin B. Korb, Carlo Kopp and Lloyd Allison (1997) A Statement on Higher Education Policy in Australia. Dept Computer Science, Monash University, Melbourne, 1997. This is our submission to the West Higher Education Review Committee.

Kevin B. Korb (1998) Research Writing in Computer Science. This is an updated (1998) version of: Technical Report 97/308. Dept Computer Science, Monash University, Melbourne, 1997. This explains some of what goes into good research writing, including argument analysis and an understanding of cognitive errors that people are prone to make. It also discusses research ethics.

Current Research Projects and Publications

For a complete list of my publications since 1993 see the publication page at Monash University.

  • Causal Discovery

    The aim of this project is to develop methods for automating the learning of causal structure from observational and experimental data. This has become a very hot topic in the artificial intelligence and statistics communities during the 1990′s as the importance of graphical representations (Bayesian networks, causal models) of probabilistic reasoning for AI has become clear, and so the need for automating their learning grows. This project has generated a number of programs which learn causal models from observational data, using greedy search, genetic algorithms and stochastic sampling to search the space of causal models, and a Minimum Message Length (MML) encoding to weigh them by posterior probability. The programs can discover networks with discrete or continuous variables.


    • Bayesian Artificial Intelligence (2004) (with Ann Nicholson) is our textbook on knowledge engineering and data mining with Bayesian networks, published by CRC Press. See our BAI book page for material supplementing the book, including illustrative source code, networks for use with problems and an updated appendix reporting Bayesian net and causal discovery tools.

    • The Power of Intervention (with Erik Nyberg). Forthcoming in Minds and Machines. We present a mathematical theory of causal intervention in linear models, demonstrating that, while causal faithfulness or simplicity may be undesirable in special cases, their counterparts in the augmentation (intervention) space are desirable. We prove that (in somewhat idealized circumstances) interventions have the capability of entirely eliminating models alternative to the truth.

    • An Information-theoretic Approach to Causal Power (with Luke Hope). Technical Report 2005/176. We are developing a new metric for assessing causal power in Bayesian networks. Unlike the metrics of Glymour, Cheng and Hiddleston, ours allows for the full representational power of Bayesian networks, including cases of intransitive causality. A shortened version was published in the Australasian AI’2005 conference.

    • Varieties of Causal Intervention (with Lucas R. Hope, Ann E. Nicholson and Karl Axnick). PRICAI 2004. The use of Bayesian networks for modeling causal systems has achieved widespread recognition with Judea Pearl’s Causality (2000). There, Pearl developed his do-calculus for reasoning about the effects of deterministic causal interventions on a system. Here we discuss some of the different kinds of intervention that arise when indeterminstic interventions are allowed, generalizing Pearl’s account. We also point out the danger of the naive use of Bayesian networks for causal reasoning, which can lead to the mis-estimation of causal effects. We illustrate these ideas with a graphical user interface we have developed for causal modeling.

  • Probabilistic Causality

    The most plausible understanding of the probabilities in causal Bayesian networks is in terms of the propensity interpretation (see, e.g., recent work by Donald Gillies). On that basis it is possible to start making sense of type and token causal relations in reference to Bayesian networks.


    • A Criterion of Probabilistic Causality (with Charles Twardy). Philosophy of Science, 2004. The investigation of probabilistic causality has been plagued by a variety of misconceptions and misunderstandings. One has been the thought that the aim of the probabilistic account of causality is the reduction of causal claims to probabilistic claims. Nancy Cartwright (1979) has clearly rebutted that idea. Another ill-conceived idea continues to haunt the debate, namely the idea that contextual unanimity can do the work of objective homogeneity. It cannot. We argue that only objective homogeneity in combination with a causal interpretation of Bayesian networks can provide the desired criterion of probabilistic causality.

    • Causal Reasoning with Causal Models (with Charles Twardy, Toby Handfield and Graham Oppy). Technical Report 2005/183; under submission to Synthese. We introduce and discuss the use of Bayesian networks for causal modeling. Despite their growing popularity and utility in this application, numerous objections to it have been raised. We address the claims that Chickering’s arc reversal rule undermines a causal interpretation and that failures of Reichenbach’s Common Cause Principle, or again failures of faithfulness, invalidate causal modeling. We also argue against Pearl’s deterministic interpretation of causal models. Against these objections we propose new model-building principles which evade some of the difficulties, and we put forward a concept of causal faithfulness which holds when faithfulness simpliciter fails. Finally, we particularize our account of type causal relevance to token causal relevance, providing an alternative to the recent deterministic accounts of token causation due to Hitchcock and Halpern and Pearl.

  • Evaluation Theory

    Recently Luke Hope and I have been investigating means of evaluating machine learning algorithms when cost-sensitive classification is not an option, using the concept of information reward. I am currently investigating also improved methods for assessing causal discovery algorithms in particular.


    • A Bayesian Metric for Evaluating Machine Learning Algorithms (with Luke Hope). The Australasian AI Conference, 2004. How to assess the performance of machine learning algorithms is a problem of increasing interest and urgency as the data mining application of myriad algorithms grows. The standard approach of employing predictive accuracy has, we argue rightly, been losing favor in the AI community. The alternative of cost-sensitive metrics provides a far better approach, given the availability of useful cost functions. For situations where no useful cost function can be found we need other alternatives to predictive accuracy. We propose that information-theoretic reward functions be applied. The first such proposal for assessing specifically machine learning algorithms was made by Kononenko and Bratko (1991). Here we improve upon our earlier Bayesian metric (Hope and Korb, 2002), which provides a fair betting assessment of any machine learner. We include an empirical analysis of various Bayesian classification learners, ranging from Naive Bayes learners to causal discovery algorithms.

  • Informal Logic and Argumentation

    Our NAG project produced the first computational model of argumentation to employ Bayesian networks to model inductive and uncertain reasoning in argument generation and analysis. Two distinct networks are employed, one for modeling the cognition of the human user and one for modeling normative reasoning in the domain. The human cognitive model is the first to incorporate the statistical illusions widely studied by cognitive psychologists with a computational model of Bayesian reasoning.

    I am currently working on applying Bayesian principles to improve argument analysis.


    • Bayesian Informal Logic and Fallacy. Informal Logic, 2005. Bayesian reasoning has been applied formally to statistical inference, machine learning and analysing scientific method. Here I apply it informally to more common forms of inference, namely natural language arguments. I analyse a variety of traditional fallacies, deductive, inductive and causal, and find more merit in them than is generally acknowledged. Bayesian principles provide a framework for understanding ordinary arguments which is well worth developing.

    • A Cognitive Model of Argumentation (with Richard McConachy and Ingrid Zukerman). Cognitive Science, 1997, Stanford. In order to argue effectively one must have a grasp of both the normative strength of the inferences that come into play and the effect that the proposed inferences will have on the audience. In this paper we describe a program, NAG (Nice Argument Generator), that attempts to generate arguments that are both persuasive and correct. To do so NAG incorporates two models: a normative model, for judging the normative correctness of an argument, and a user model, for judging the persuasive effect of the same argument upon the user. The user model incorporates some of the common errors humans make when reasoning. In order to limit the scope of its reasoning during argument evaluation and generation NAG explicitly simulates attentional processes in both the user and the normative models.

  • Artificial Evolution

    This project applies artificial life simulation techniques to issues arising in evolution theory and evolutionary psychology.


    • The Evolution of Aging (with Owen Woodberry and Ann Nicholson). Australasian ALife 2005. Based on the early group selection model of Gilpin (1975) for the evolution of predatory restraint, Mitteldorf (2004) designed an ALife simulation that models the evolution of aging and population regulation. Mitteldorf sees the evolution of aging as a case of extreme altruism “in the sense that the cost to the individual is high and direct, while the benefit to the population is far too diffuse to be accounted for by kin selection.” We demonstrate that Mitteldorf’s simulation is dependent on kin selection, by reproducing his ALife simulations and then introducing a mechanism to remove all and only the effects of kin selection within it. The result is the collapse of group selection in the simulation, suggesting a new understanding of the relation between group and kin selection is needed.

    • An ALife Investigation of the Origins of Dimorphic Parental Investments (with Steven Mascaro and Ann Nicholson). Australasian ALife 2005. When Trivers (1972) introduced the concept of parental investment to evolutionary theory, he clarified many of the issues surrounding sexual selection. In particular, he demonstrated how sex differences in parental investment can explain how sexually dimorphic structure and behaviour develops in a species. However, the origins of dimorphic parental investments also need explanation. Trivers and others have suggested several hypotheses, including ones based on prior investment, desertion, paternal uncertainty, association with the offspring and chance dimorphism. In this paper, we explore these hypotheses within the setting of an ALife simulation. We find support for all these alternatives, barring the prior investment hypothesis.

  • Philosophy of AI

    Are artificial intelligences possible? What kinds of design might they take? Can logical reasoning suffice for an intelligence? What to think of McCarthy and Hayes’s Frame Problem? What is thinking? What are minds? And other ponderables…


    • Machine Learning as Philosophy of Science. Minds and Machines, 2004. I consider three aspects in which machine learning and philosophy of science can illuminate each other: methodology, inductive simplicity and theoretical terms. I examine the relations between the two subjects and conclude by claiming these relations to be very close.

    • Inductive Learning and Defeasible Inference. Journal of Experimental and Theoretical AI, 1995. The symbolic approach to artificial intelligence research has dominated AI until recent times. It continues to dominate work in the areas of inference and reasoning in artificial systems. I argue, however, that non-quantitative methods are inherently insufficient for supporting inductive inference. In particular there are reasons to believe that purely deductive techniques—as advocated by the naive physics community—and their nonmonotonic progeny are insufficient for supplying means for the development of the autonomous intelligence that AI has as its primary goal. The lottery paradox points to fundamental difficulties for any such non-quantitative approach to AI. I suggest that a hybrid system employing both quantitative and non-quantitative modes of reasoning is the most promising avenue for developing an intelligence that can avoid both the paralysis induced by computational complexity and the inductive paralysis to which purely symbolic approaches succumb.

    • Symbolicism and Connectionism: AI Back at a Join Point, Information, Statistics and Induction in Science, ISIS 1996. Artificial intelligence has always been a controversial field of research, assaulted from without by philosophers disputing its possibility and riven within by divisions of ideology and method. Recently the re-emergence of neural network models of cognition has combined with the criticisms of Hubert Dreyfus to challenge the pre-eminent position of symbolicist ideology and method within artificial intelligence. Although the conceits of symbolicism are well worth exposing, the marriage between connectionism and Dreyfus’s philosophical viewpoint is unnatural and much of the disputation between connectionism and “traditional” artificial intelligence misbegotten.

    • The Frame Problem: An AI Fairy Story. Minds and Machines, 1998. I analyze the frame problem and its relation to other epistemological problems for artificial intelligence, such as the problem of induction, the qualification problem and the “general” AI problem. I dispute the claim that extensions to logic (default logic and circumscriptive logic) will ever offer a viable way out of the problem. In the discussion it will become clear that the original frame problem is really a fairy tale: as originally presented, and as tools for its solution are circumscribed by Pat Hayes, the problem is entertaining, but incapable of resolution. The solution to the frame problem becomes available, and even apparent, when we remove artificial restrictions on its treatment and understand the interrelation between the frame problem and the many other problems for artificial epistemology. I present the solution to the frame problem: an adequate theory and method for the machine induction of causal structure.