Biology in machine intelligence: Michie’s ‘optimisable organisms’

Blog series Science

Author Aswin Valsala Narayanan, WRoCAH Collaborative Doctoral Award student researching the Donald Michie Archive

What does biology have to do with artificial intelligence (AI)? Quite a bit actually, considering that the whole field of AI began with the question of whether it would be possible to mechanise the nervous system and create machines that could simulate ‘intelligence.’ While this led the early pioneers to study how the nervous system and the brain worked, their research was largely focussed on understanding how to mimic human cognition, that is, the learning processes. Most of these pioneers were mathematicians, psychologists, or engineers, whose interest in biology was restricted or defined by their goals to mechanise intellect.

Donald Michie (1923-2007), the person who almost single-handedly established AI research in Britain, was the exception. Although Michie’s interest in AI began during the Second World War, as a teenage codebreaker at Bletchley Park who befriended Alan Turing, he studied medicine after the war and eventually became a geneticist.

Hence his interest in biology was not just about mimicking cognition to develop what he preferred to call ‘machine intelligence’. Unlike Turing, or the other early American AI pioneers such as John McCarthy, Norbert Weiner, Claude Shannon, and Marvin Minsky, Michie was a biologist by training, and had a remarkable career as a maverick geneticist during the 1950s. Along with his then-life and research partner Anne McLaren, he was involved in pathbreaking embryology research that laid foundations for the development of in-vitro fertilisation (IVF) and later co-wrote one of the first textbooks on molecular biology.

Michie’s focus shifted from biology to AI research in the early 1960s while he was a Reader in the surgical sciences department of Edinburgh Medical School, and he went on to establish the first university department for machine intelligence in the world at the University of Edinburgh in 1966.

However, despite leaving biological research for AI, Michie the biologist continued to live within Michie the machine intelligence researcher. His strong background in biology was a boon as a scientist trying to create ‘learning’ machines. Michie’s first machine learning project called MENACE (Machine Educable Noughts and Crosses Engine) was in fact a simple yet brilliant simulation of reinforcement learning, a process by which animals learn adaptations to an environment through trial and error. It was a mechanical computer, made of matchboxes with coloured glass beads inside, that was capable of learning to play a perfect game of noughts and crosses (tic-tac-toe). Each matchbox of MENACE corresponded to a game state, with the glass beads inside representing which move to make if that state was reached during a game. If a game was lost, then the beads used to decide the moves were removed from their respective boxes, thus reducing the chance of making those moves in future games. However, if the game was won, then extra beads of the same kind that were picked from the boxes were added to respective boxes, thus reinforcing the chances of making an ideal move.

MENACE was not just a practical breakthrough by a budding machine intelligence researcher with a keen understanding of biology, psychology and learning machines. It was also a geneticist’s exploration of ‘learning’ as a Darwinian evolutionary process. In the conclusion of the chapter ‘Trial and Error’ for the Penguin Science Survey 1961 Part 2, edited by S. A. Barnett and McLaren, Michie recommends how his matchbox model can be thought of as a ‘Darwinian’ system:

‘…thinking of the boxes as discrete habitats through which is dispersed a species of organism which takes the form of glass beads. Each habitat is occupied by several varieties, one of which, by reason of greater fitness to that particular habitat, usually displaces the other and becomes the local type.’

He expanded this biological lens further in a chapter on the BOXES learning algorithm (developed based on his MENACE matchbox model) that he co-wrote with R. A. Chambers for biologist C.H. Waddington’s 1968 book The Origin of Life. The chapter (titled ‘BOXES as a model for pattern formation’) was about how Michie and Chambers attempted to solve a difficult problem in the adaptive control of a physical apparatus, the properties of which are at first more or less unknown. Their aim was to create a system that is self-designed, utilising its trials and failures as the only source of information to learn how to control the apparatus:

‘The computer program which was written to perform this task thus exemplifies a learning’, or ‘evolutionary’, or ‘adaptive’, or ‘self-organizing’ system… In conclusion the adaptive system described possesses features in common with various self-organizing biological systems, notably evolution, morphogenesis and learning.’

Michie predicted that future ‘machine builders’ will be forced to understand in increasing depth and detail the logical and mathematical structure of learning processes. However, he didn’t see this as some one-way conceptual pipeline from biology to machine intelligence. Michie believed that biologists too can gain from machine intelligence research:

‘The biologist will be able to use the results of these analyses to sharpen his investigation of living nervous systems, which quite possibly operate through entirely different mechanisms. At the same time, whenever a learning machine exhibits a striking parallel with human or animal behaviour, the biologist should be on the alert: it may be a clue to a biological mechanism.’

But he also cautions biologists against reading too much from a model or a device capable of performing a bodily function. Although a model of a biological function can be illuminating, Michie warns against the temptation to think that the way the device works must be how the body works too. He points out that all we can expect from a model is that it may deepen our understanding ‘of the matrix of physical laws within which both the model and the biological system have to work,’ much like how the study of aeroplane flight can help us understand the aerodynamics laws to which flying beings are also subject to.

Understanding of biology and experience in biological research helped Michie a great deal in setting up the first AI research centre outside of the United States at the University of Edinburgh. In 1963, Michie’s Experimental Programming Unit (EPU) developed as a special project under the umbrella of the Edinburgh Medical School that he was a part of. Later, when the Department for Machine Intelligence and Perception was established in 1966, Michie’s biological expertise was put to great use in leading the team that built the FREDDY II robot. FREDDY II, dubbed by National Museums Scotland as the world’s first ‘thinking robot’, combined a seeing eye and feeling hand.

Michie also applied his information science and machine intelligence perspectives to biological problems. His 1977 letter to biologist Sydney Brenner, written presumably after he heard Brenner speaking about the relationship between the length of the DNA and its computational capability, is a great example of how Michie applied a computational lens to a biological problem. In the letter, Michie seems to be extremely impressed by Brenner’s work and suggests a way of conceptualising the process of ‘making’ an organism in terms of computation capability:

‘Let me take you up on a small point in connexion with the relative lengths of the DNA codescript of E. coli, Drosophila, mouse and man. What you say concerning the correlation between DNA- quantity and the complexity of the finished organism is OK in a rough and ready sort of way. But to do it properly one must remember that the whole point of biological functions is that they must be computable by the intracellular apparatus within certain time-limits, and that the genetic programs of the above organisms have very different running times in which to generate the finished article - hours for E. coli, days for Drosophila, weeks for mouse and years for man.’

Michie explains this further using the terminology that he developed for his 1977 paper ‘Practical limits to computation’, which he enclosed with the letter to Brenner. Perhaps the most pertinent insight into how Michie applied his understanding of biology to his AI research, and vice versa, comes in the final part of the letter where he quips how this ‘speculation’ on how DNA computes information can be studied better ‘after one has actually made, tested and proved things about at least one respectable ‘informational organism’.’ He further connects this to his chess end-games research quite organically:

‘I regard my chess end-game programs as optimisable organisms, and with Bratko have just succeeded in putting together a simple but respectable unicellular one (i.e. consisting of a single ‘Advice Table’, a semi-isolated package of locally relevant heuristic knowledge). It synthesises strategies on demand for the correct defence of King and Knight against King and Rook.’

Michie goes on to suggest that he was planning an even more complicated end-game program that would require incorporating ‘inductive learning technique’, something that he would use to challenge the researchers and students he mentored during the late 70s and early 80s. These end-game problems eventually became crucial to the inductive learning tools and methods developed by researchers like J. Ross Quinlan, Alan Shapiro, Tim Niblett and Stephen Muggleton. Quinlan’s pathbreaking iterative dichotomiser 3 (ID3) learning algorithm for generating decision trees was a product of an end-game problem that Michie posed to him.

The final short paragraph in the letter to Brenner provides a succinct insight into Michie’s use of biology in his AI research:

‘I never mention biology when presenting this stuff since I am not logically required to do so. But I have tried to make use of being a biologist in developing this approach to complexity.’

Further reading

The Donald Michie Papers at the British Library comprise two separate tranches of material gifted to the Library in 2004 and 2008. They contain correspondence, notes, notebooks, offprints and photographs and are available to researchers.

Muggleton, Stephen H. ‘Inductive Acquisition of Expert Knowledge’, 1986.

Quinlan, J. R., ‘Induction of Decision Trees’, Machine Learning 1, no. 1 (March 1, 1986), 81–106

Quinlan, J. R., ‘Discovering Rules by Induction from Large Collections of Examples’, Expert Systems in the Micro Electronics Age, 1979

Shapiro, A. and Niblett, T. ‘Automatic Induction of Classification Rules for a Chess Endgame’, in Advances in Computer Chess, ed. M. R. B. Clarke, Pergamon Chess Series (Pergamon, 1982), 73–92.