In observing artificial intelligence systems, it appears that there are six pri-mary sources of inspiration to the problem of designing artificial intelligence systems: Introspection, computational theory, statistics, game theory, biol-ogy & neuroscience and psycholbiol-ogy. There are upsides and downsides to using every one of these sources of inspiration, and different people’s weightings of each of those upsides and downsides could lead to a different conclusion as to which source of inspiration is likely to lead to the best results.
Inspiration from introspection refers to the approach whereby a system de-signer designs their system based on internal observation of their own thought processes, relying on the fact that they are themselves an example of what they seek to build. Some early research in artificial intelligence appears to be of this sort, and was discussed by McCarthy & Hayes:
“Programs have been written to solve a class of problems that give humans intellectual difficulty. . . . In the course of designing these programs intellectual mechanisms of greater or lesser generality are identified sometimes by introspection, sometimes by mathemati-cal analysis, and sometimes by experiments with human subjects.
Testing the programs sometimes leads to better understanding of the intellectual mechanisms and the identification of new ones.”
(McCarthy & Hayes, 1969, pp. 465–466) In one guise or another, introspection has been long associated with the study of human thought and can be traced to Socrates (Schultz & Schultz, 1996, p. 77). The approach was used in a rigorous way in the early days of the foundation of psychology as an independent discipline and had one of the discipline’s founders, Wundt, as a proponent (Schultz & Schultz, 1996, p. 77).
The central argument for the use of introspection is that due to the subjective
nature of conscious experience, the only person able to observe it is the person who is having the experience. In terms of the design of artificial intelligence systems, the approach also requires little background knowledge and so allows for quick initial progress, this could be a reason for its early use within artificial intelligence. Within psychology, the approach is now largely seen as lacking rigor. The best known argument against introspection is by Watson (1913), whose main argument is that due to the subjectivity of introspection, the results gained by using it cannot be reliably confirmed. Watson argued that one should only observe behaviour and not attempt to theorise on internal mental states – a school of thought that became known as behaviourism.
In more recent times, introspection in psychology is deemed unreliable as it requires a-priori knowledge of one’s own unconscious thought process. It has been shown by psychological experimentation that it is not possible to know this (Nisbett & Wilson, 1977).
Computational-theoretical, statistical and game-theoretical inspiration ap-pears to have similar upsides and downsides and so shall be discussed together.
Each of these inspirations holds its basis in mathematical proof. This leads to the design of systems that can be seen to be highly rational, sometimes having provably perfect rationality. Their downside, however, is that the computa-tional resources such systems require can make them infeasible to use in all but the smallest problem instances. This issue leads to the problem of marshalling computational resources to optimise between rationality of an action and the reaction time (Russell et al., 1993).
Using biology and neuroscience for inspiration does imply at least some level of in-built optimality between rationality and response time, as arguably evolutionary forces have already had to solve the issue (though whether those optimisations can apply to any computational version remains to be seen). A benefit of using neuroscience as a source of inspiration is that it is a source of ideas for systems that have not been originated by another human. This means that arguably it provides ideas that may not occur to a human approaching the problem of artificial intelligence from either an introspective or mathematical basis. The drawback of using neuroscience or biology as a source of inspiration is that the level of abstraction is too low to easily derive intelligent behaviour.
To use the well-worn brain-computer analogy, the task of neuroscience is anal-ogous to attempting to reverse-engineer a database server by analysing the function of each individual transistor. It is theoretically possible, but would take a great deal of effort. This is not to say that the work of neuroscience should not be done, it provides a highly valuable contribution, however this thesis holds that using the field as a primary source of inspiration for artificial intelligence systems will only yield advances at a very slow rate. A better role would be to indirectly influence artificial intelligence by providing a basis and plausibility checks for psychological theories.
Designing artificial intelligence systems by drawing inspiration from psy-chology shares the same advantages, to differing extents, as that of biology and
neuroscience inspired systems, but does not have the drawback of an abstrac-tion level that is too low to be of direct use. There are other, weaker drawbacks though. Psychological theories on the whole are vaguer than those of other approaches. This however can be turned into a positive point as it allows an artificial intelligence system designer enough leeway to implement the system in a manner that suits the mostly serial nature of computer processing. An analogy would be that psychological theory can be used as a form of software engineering specification. Another drawback is that due to not being able to properly control all experimental variables, data from experimental psychol-ogy is inherently noisier than the physical sciences. This leads to a greater variety of psychological theories that fit the data, but contradict one another, leading to a higher likelihood of a theory utilised by the artificial intelligence system designer being incorrect. Poole, Mackworth & Goebel (1998) make an argument against taking inspiration from psychology, neuroscience or biology.
The argument is one by analogy with the development of flight:
“First note that there are several ways to understand flying. One is to dissect known flying animals and hypothesize their common structural features as necessary fundamental characteristics of any flying agent. With this method an examination of birds, bats, and insects would suggest that flying involves the flapping of wings made of some structure covered with feathers or a membrane. . . . An al-ternate methodology is to try to understand the principles of flying without restricting ourselves to natural occurrences of flying. This typically involves the construction of artefacts that embody the hy-pothesized principles, even if they do not behave like flying animals in any way except flying. This second method has provided both useful tools, airplanes, and a better understanding of the princi-ples underlying flying, namely aerodynamics.”
(Poole et al., 1998, pp. 2–3) This argument ignores the fact that it was only through the use of ob-servations of biological flight that the principles of aerodynamics were discov-ered. Sir George Cayley, in publishing his findings on aerodynamics, effectively founding the field, made the observation that birds do not flap once full ve-locity has been reached (Cayley, 1809, p. 167). Poole et al.’s argument is that by only using examples from nature, it is restricting the space of intelli-gent entities, and this hampers progress in the development of the underlying principles. With the analogy, it is quite obvious that aeroplanes and birds use different methods of flight but both use the same underlying principles.
However, the argument that the principles of aerodynamics came about due to some elementary derivation is fallacious, as was described earlier. As such, Poole et al.’s argument that such an approach should also apply to artificial intelligence is precarious – we don’t yet know enough about the principles of intelligence to use them without reference to existing intelligent beings.
Due to the weighting placed on the advantages and disadvantages of each of these sources of inspiration, this thesis uses psychology. Ultimately, all sources of inspiration have a place in contributing towards the problem of artificial intelligence, and any full solution will be a synthesis of ideas from all fields, but from the point of view of a single project, there will always be one field that is used more than the others.