Abductive Inference: Computation, Philosophy, Technology

In the first article, abduction is defined as a form of inference that starts with data that describes something and formulates a hypothesis that gives the best explanation of the data. The term abduction is credited to the philosopher/logician Charles Sanders Peirce, and has been described by the author "modus ponens" turned backward. The author gives examples of abductive reasoning from everyday life and science, and he is careful to note that problems with combinatorial explosion that results from generating all possible explanations, and so he takes abduction to include generation, criticism, and possible acceptance of explanatory hypotheses. Abductions are described as "ampliative" inferences, in that there is an increase in information after accepting a best explanation, i.e. successful abductions are "truth producing". This is actually a very controversial claim, since it is claiming that the abductive process is creative. This is not surprising from a human reasoning point of view, since humans clearly exhibit creativity. But from a machine point of view if abduction can be implemented on a machine, then so can creativity, and the latter claim is currently hotly debated.

The next article discusses abductive reasoning in the context of the most popular knowledge-based systems. The authors address the current status of AI as being one that is fragmented into several competing paradigms, and point to the strong disagreements on how to quantify progress in AI. They argue for the need for AI to continue as a pluralistic enterprise with opposing viewpoints and ideas, and list four different ways in which AI has viewed the idea of a program. One of these views is the "strong AI" viewpoint, which has sometimes dominated the popular view of AI, and the authors argue correctly that whether or not machines can have humanlike cognitive states is completely irrelevant to the use of AI technology. This they call "AI as design science" which attempts to view intelligent agents in the abstract, with humans being one set, and silicon-based machines another. This view of AI is the one that is the most effective from the standpoint of business and industrial applications, and this is due to its insistence on practical application and the consequent minimization of philosophical debate, the latter of which one can argue has crippled progress in AI, or at least delayed it considerably. The goal of the design paradigm for AI, as they authors explain it, is to find the general principles of computation and information processing that subsume the human case. After all, they argue, human thinking is of a "black box" nature, as we currently understand it (our understanding though increasing dramatically with every passing day). The reasoning systems used by machines can be studied, understood, and altered as we please, and in fact could teach humans how to reason better. They also discuss the differences between the symbolic and connectionist approaches to AI, arguing that both of these should be subsumed into a more abstract level of description, called the "information-processing" level. The input, output, and types of information processing are viewed as the "top-level" content for constructing theories of AI, two of these being the symbolic and connectionist viewpoints. The authors clearly believe that the logical tradition in AI has caused difficulties, in that it has separated knowledge from its functions, and this leads to the omission of important aspects in knowledge representation. This motivated the "frame approach" to knowledge representation, and the authors discuss three different reasons for using this representation: 1. Its utility and efficiency in organizing knowledge about classes of objects. 2. The ability to create type-subtype hierarchies. 3. The possibility of embedding procedures in frames to allow inferencing. Object-oriented programming is mentioned as sharing much in common with the frame approach. The authors then finally discuss applications, such as diagnostic reasoning, which from a commercial standpoint has proven to be a very useful application of abductive reasoning. The computational complexity of the diagnostic problem is pointed out, illustrating the need for heuristics in the obtaining of a solution in a reasonable time frame. Most interesting in this discussion is that the authors ask what kind of intelligence is needed to perform diagnostic reasoning. They make a connection here with some current research that attempts to define intelligence independent of what is done in the human case. The answer of how diagnosis is to be done needs to be answered in the context of "generic mental structures." A different mental structure will give a different answer, they argue. In the authors view, diagnostic reasoning may involve "malfunction hierarchies", "rule-out" strategies, etc, and so one needs approaches that directly address the higher level issues of knowledge-based reasoning. As a science of intelligence therefore, the task of artificial intelligence should be to identify concretely the strategies for processing information and their coherence.