"If the field of artificial intelligence had concentrated on building artificial footballers rather than artificial chess players, how might our efforts to build artificial intelligences be different today?"
This essay intends to address the issue of what was, what is and what might have been in Artificial Intelligence (AI) research, and how that might relate to the production of human-like machines in the future. Exploration of this theme will cover an overview of the current status of AI, the past developments in chess-like domains, the historical and more recent developments in football-like domains, and what is still left to be done. The intention is to provide an insight into a possible world where the last fifty years of AI research saw more informal, behaviour-based methods take the front line, and the formal "good old fashioned AI" (GOFAI) (1) techniques take a back seat. I hope to show that this world would have been a subtly and fascinatingly different one had psychologists, philosophers and biologists not parted company with AI for most of the latter half of the twentieth century.
The current state of AI covers several areas - computer vision, artificial life, speech recognition, logic, representation and reasoning systems, bio-inspired systems, and more - which wax and wane in popularity as time progresses. The pioneers of modern AI however were interested in all of these fields much more equally. Alan Turing, for instance, not only gave us the Turing Test (2) and Turing Machines (3) but was also later interested in areas in the natural world such as the mathematics behind plant growth. The post war field of cybernetics also produced many results and ideas which are only now being rediscovered and coming back into practice. The work of the late 1960's to the mid 1980's though are what recent pioneers like Cliff (4) and Brooks consider to be the GOFAI areas which dominated the direction and focus of research in the field. These areas are best categorised by Newell and Simon (5), who first offered a formalised manner in which to apply research in computer science to help understand real minds. However, some people came to consider 'Symbols and Search' to be somewhat stagnant and AI in general to be in need of revolution:
"Artificial Intelligence research has foundered in a sea of incrementalism." Rod Brooks, (6).
Concerning the issue of why chess is considered an AI problem, we must again turn to Turing. In his paper 'Computing Machinery and Intelligence' (2), Turing describes some motivations for creating an artificial chess player (more precisely, a game playing machine that aims to simulate the responses of a disembodied human being) . Some of the points Turing makes are addressing the practical engineering limitations of the time; for instance he points out that "No engineer or chemist claims to be able to produce a material which is indistinguishable from the human skin". However, with hindsight it is possible to see that Turing, just like Newell and Simon after him, may have inadvertently steered AI away from its multidisciplinary orientation into the formal, logical discipline which may have stifled it later on.
Clearly for a researcher, chess as an AI task is at once stimulating, understandable, achievable and can have demonstrable success. Furthermore, it is a discretised task in a closed domain with all information about the environment available to the player. As something which does not require a physical presence it is one of the few tasks suited to a top-down approach to the solution. The top-down approach as described by Brooks (7) is that where higher level aspects of intelligent behaviour (such as symbol manipulation) are modelled first, within frameworks such as the sense-model-plan-act (SMPA) framework. This contrasts with bottom-up techniques, more descriptively considered to be behaviour based, where complex behaviour is designed in through simple rules at a basic level (this is similar to, but not quite, a sense-act framework since this implies complete reactivity).
The move towards more human-like artificial intelligences, perhaps even those that could take part in a game of football, is the natural progression from the robot-based work of Brooks, Cliff et al. Just like Turing did, I would like to avoid the engineering complications in my discussion, but rather than choose a domain where embodiment is irrelevant. I will simply say that I am discussing the possibility that if the body of a footballer (artificial or otherwise) was available with which to interface an artificial brain, could we produce behaviour both appropriate and convincing such that it could be released as a fully autonomous agent into a real game of football? Such a problem could be solved in simulation, and the results could await a sufficiently complex machine to complete the embodiment (compare how chess researchers needed to wait for sufficient computational power for their search algorithms to be able to beat a Grand Master). In the possible world I mentioned earlier, perhaps now we would be at this stage, awaiting advances in whatever technologies might be appropriate such as cloning or bio-mechatronics.
The GOFAI techniques, suited to chess though they may be, could still offer us a head start in our quest for a footballer's brain. For instance, though these techniques may be next-to useless when attempting to distinguish near and far or fast and slow, they may be ideally suited to formulating tactics and set-pieces. Upon analysis of what it is that footballers actually do, it seems our primary aims would be sensory-motor co-ordination and teamwork (while secondary aims for full realism could include showmanship, excitement and passion). Concentrating on the primary objectives, it seems sensory-motor co-ordination is ideally suited to the bottom up approach favoured by Brooks, whereby one 'layer' at a time we could produce the abilities required in our footballer. I believe however that teamwork, and with it tactics and set-pieces, might be better suited to a 'Symbols and Search' approach. Although it can be shown that methods such as dynamical neural networks (8) offer a way to represent simple states in continuous systems, they tend to be used in situations where explicit planning is not required in the way I think it is for football.
If we take Beer's broad definition of an autonomous agent as:
"any embodied system designed to satisfy internal and external goals by its own actions while in continuous long-term interaction with the environment in which it is situated." Beer, (8).
Then clearly an artificial footballer falls into this category since she must use her own actions during 90 minutes of continuous interaction with the other players on the pitch to win the game by satisfying the external goals (pun intended). As such, the footballer must learn to cope with its surroundings without support. Bio-inspired techniques offer the possibility of just such a robust solution, with the added advantage of graceful degradation. That is, we want a bio-inspired design for the controller of our footballer, so that in the event of damage caused to either the machine or the control unit, as much functionality as possible is retained. Old fashioned AI techniques would fail at this point, for example if the section for processing sensor data into a world model was lost, then a blocksworld implementation would not be able to plan or act.
Finally, I wish to directly address the question, "If the field of AI had concentrated on building artificial footballers rather than artificial chess players, how might our efforts to build AI's be different today?" by stating a belief that both the new behaviour-based, bottom-up approach to AI, and what Brooks calls GOFAI techniques will both have their uses in the artificial footballers of the future. Had AI research gone the other way in the late fifties and concentrated on cybernetics and robotics, I think we would now be seeing a mathematical, logical, backlash against those techniques. It has taken us fifty years to get close to perfecting 'Symbols and Search' and this year's attempts to pass the Turing Test only just managed to fool half of the judges some of the time. Bio-inspired computing will come to even greater prominence as the need arises to simulate things we don't fully understand. Whilst we play catch-up with the possible world where those techniques are already prominent, so that possible world will only just be discovering the knowledge we already have of those "good old fashioned AI" techniques.
- Rod Brooks (1997). From earwigs to humans, Robotics and Autonomous Systems, 20:291-304
- Alan Turing (1950). Computing Machinery and Intelligence, Mind, 49:433-460
- Alan Turing (1937). On Computable Numbers, with an application to the Entscheidungsproblem, Proc. Lond. Math. Soc. (2) 42:230-265 (1936-7); correction ibid. 43:544-546.
- Dave Cliff (1994). AI and A-Life: Never mind the Blocksworld. In Tony Cohn (ed.) Proceedings of the 11th European Conference on Artificial Intelligence, pp 803-808. John Wiley and Sons.
- Allen Newell and Herbert Simon (1976). Computer science as empirical inquiry: Symbols and Search. Reprinted in John Haugeland (ed. 1997) Mind Design II, pp. 81-110. MIT Press.
- Rod Brooks (1990) Elephants don't play chess, Robotics and Autonomous Systems 6:3-15.
- Rod Brooks (1995). Intelligence without reason. In Luc Steels and Rod Brooks (eds.) The Artificial Life Route to Artificial Intelligence, pp. 25-81 Lawrence Erlbaum Associates.
- Randy Beer (1995) A dynamical systems perspective on agent-environment interaction, Artificial Intelligence, 72:173-215.
This was my attempt at assignment 1 for the level 3 computing module Bio-Inspired Computing at the University of Leeds. Rushed as usual, but written with good intentions... I'll post the mark when I get it.
News just in... the mark was OK (54/75 I think) and the feedback said it was a good essay but I didn't really answer the question!