There aren’t many sci-fi movies as cleverly written and executed as the 1983 Hollywood classic WarGames, directed by John Badham and written by Lawrence Lasker and Walter F. Parkes. But it might not have been as good as it is if not for the input of the British AI scientist Donald Michie (1923-2007) and his knowledge of game-learning machines, especially computers that can play noughts and crosses.

Who was Donald Michie?

Donald Michie was a codebreaker at Bletchley Park (BP) during the second world war; after the war, a geneticist and later an artificial intelligence (AI) pioneer who founded the first department for machine intelligence in the world at the University of Edinburgh. He made significant contributions to all three fields that he worked in and was instrumental in institutionalising AI and machine learning (ML) research in the UK, and abroad.

A major part of his AI research focused on machines that could play games. However, one of his most famous achievements in this subfield of AI is MENACE (Machine Educable Noughts and Crosses Engine), a mechanical computer made of matchboxes and beads that could learn to play noughts and crosses. MENACE is now considered one of the first working examples of, what ML researchers call, reinforcement learning. This eventually led to the development of the BOXES algorithm by Michie and R. A. Chambers for learning of dynamic systems.

Therefore, it is no surprise that Universal Studios reached out to him when working on WarGames.

Dangerous games

If you haven’t seen it, be warned that there will be spoilers ahead.

WarGames tells the story of a young computer hacker, David Lightman (Matthew Broderick) who accidentally gains access to a US military supercomputer that controls the launch of nuclear warheads. He starts playing games with it thinking that it’s a game-playing machine, which makes the supercomputer run real war-game simulations and almost launch the nuclear warheads. Fortunately, Lightman, along with Prof. Stephen Falken (John Wood), the scientist who designed the supercomputer, makes it play a game of noughts and crosses against itself, which leads the machine to conclude that a Third World War involving nuclear warheads cannot be won, thus stopping it from launching the missiles.

It might be an exaggeration to call it sci-fi as it hits a little too close to reality than most sci-fi movies would. Although there is no public record of any supercomputer malfunctioning and almost launching nuclear warheads to start a world war, there have been a few mishaps in United States alone that have almost caused nuclear detonations, and one incident in 1983 when a Soviet early warning system gave a false alarm of a missile launch by the US.

The fear of the Cold War escalating to a direct nuclear war was alive and well when WarGames was released. However, that was only a part of the reason why the movie became a box office success and a cult classic. It was the presence of an ‘intelligent’ supercomputer that could play games and treat global thermonuclear war as a game, and act on its own to potentially end the world as we know it, that made WarGames such a compelling film.

Michie’s concern

WarGames tapped into a real concern that was prevalent at that time, regarding the increasing inscrutability and influence of large computers that were being used by big companies and government organisations for various purposes. This was something that Michie too was very concerned about and he points out in his correspondence regarding the script consultancy with Universal Studios:

We are all in danger from the current emergence of super-computers inscrutable to man. Very few are yet alert to the seriousness and nature of the peril.

This is not surprising coming from Michie, who had already been discussing the social risks of AI for over a decade by this point. In fact, he was the one who organised the first global conference on the social impact of AI in 1972 at Serbelloni in Italy¹. In the years since then, Michie had raised some of his concerns regarding increasingly complex and sophisticated computers in several publications despite being a pioneer in the field. It makes sense why he was enthusiastic to collaborate on this sci-fi movie project, which was initially titled The Genius.

Main suggestions

“WarGames Correspondence with Universal Studios”, 1981. (Add MS 88958/3/54). Reproduced with permission of the estate of Donald Michie.

In his correspondence, Michie points out five pages worth of technical inaccuracies in the draft script. Most of them are small details regarding the terminologies that are spoken about and technical issues in how the technology is used in various scenes. Michie recommended a lot of minor changes, for instance using the term “printer” instead of “teleprinter,” which was no longer used among the computing community by the 1980s. However, he also pointed out major technical errors, such as that a program once terminated would not have a fragment of itself accidentally re-animated when the computer goes online as was shown in one of the scenes in the draft script.

Although some of Michie’s minor suggestions have been taken into account by the scriptwriters, we can see that most of the technical inaccuracies he pointed out are still there in the final cut of the movie. This includes the premise of Michie’s main suggestion: to consider changing the climax sequence where the machine is self-playing a 3-in-a-row noughts and crosses to “learn” that it’s a game that cannot be won. Michie recommended changing it a game called 5-IN-A-ROW (sometimes called Peggity) as it would take much longer for the computer to self-learn how to play compared to noughts and crosses, which can be learned by a fast computer within seconds. This, in his view, would make the long climax sequence where the computer is learning to play the game more realistic. Moreover, Michie pointed out that the supercomputer “would try to establish the ‘no win’ property by logical proof,” rather than brute-forcing through various playing strategies as is shown in the movie. He states that a computer as sophisticated as the one shown in the movie would not use a ‘try everything’ strategy to learn how to play a game.

Crucial contribution to the script

However, one key suggestion of Michie’s that was taken into account by the studio and the scriptwriters was the technique by which David and Prof. Falken make the computer play itself. In the draft script given to Michie, the computer “somehow” starts self-playing, presumably without a prompt. Michie suggested changing the number of players to “zero” to make the computer play itself. This is what happens in the movie.

Michie also recommends changing the way the supercomputer abruptly stops trying to guess the launch codes of the missiles, as it was in the draft script. He instead suggests that the filmmakers show the computer generalise from the self-play sequence that ‘nuclear wars are not winnable’ and apply this to terminate the war game it was playing. He recommends David or someone else instructing the computer to do this, although this was not taken into account. In the film, the computer automatically generalises from the noughts and crosses self-playing that various real-world nuclear war-game simulations all end up with no winners.

Although Michie’s main suggestion to change the 3-in-a-row game to 5-in-a-row was ignored by the filmmakers, he himself points out in the correspondence that it is not necessary to fix this as ‘there is a risk of distracting from the gripping simplicity of the ending by fussing over side-issues.’ It shows his concern for how well laypeople could receive the movie and its main theme of the futility of nuclear war, rather than focusing too much on the technical inaccuracies or inconsistencies in the script.

Aswin Valsala Narayanan

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.

Aswin Valsala Narayanan is as PhD student at the University of Leeds and the British Library. He is on an WRoCAH Collaborative Doctoral Award researching the Donald Michie Archive, exploring Michie's work as an artificial intelligence researcher in post-war Britain.

[1] Rosamund Powell, “The ‘Artificial Intelligentsia’ and Its Discontents: An Exploration of 1970s Attitudes to the ‘Social Responsibility of the Machine Intelligence Worker,’” BJHS Themes, September 19, 2023, pp.1–15, https://doi.org/10.1017/bjt.2023.8.

There is a good chance that AI research in Britain would not have evolved the way it did had Alan Turing been a great chess player. As a matter of fact, he couldn’t hold a candle to the chess masters, such as Hugh Alexander and Harry Golombek, who were in his Enigma codebreaking team at Bletchley Park (BP). As per one account, Turing once played himself into such a mess that his opponent Golombek turned the board around and tried to save the game for him.

However, Turing’s mediocrity in chess proved to be a blessing for one of the youngest codebreakers at BP, Donald Michie, who was just 18-years-old when he was recruited to the Government Code and Cypher School. Despite having no mathematical background, he proved to be a quick learner and eventually became a key member of Testery — Major Ralph Tester’s team attempting to break the ‘Fish’ codes using manual hand methods. It was during his time at Testery that he met and befriended Turing, who had developed Turingery — one of the manual methods used by codebreakers of Testery.

Donald Michie c. 1974 (Add MS 89072/1/5). Reproduced with permission of the estate of Donald Michie.

While Turing was no match for his chessmaster teammates, in young Michie, he found a partner who could give him a competitive game. They met every week at a pub in Wolverton not too far away from BP. This friendship and weekly chess session proved to be life-changing for Michie.

According to Michie, their conversations invariably centred around ideas of learning machines and mechanising chess, ideas formative to his interest and eventual career in machine intelligence.

Turing and Michie remained friends after the war.

Post-war games and game-playing machines

In the period of 1947-48, Michie and the mathematician Shaun Wylie, another colleague at BP, developed a chess-playing algorithm, or a machine on paper — a ‘paper machine’, so to speak. This was basically a collection of machine-rules to decide the next move using the opponent’s move as input. The ‘MACHIAVELLI’ (so named after creators Michie and Wylie), was created independently of the paper-machine that Turing and his friend David Champernowne developed, named ‘TUROCHAMP.’

I. J. ‘Jack’ Good, who was friends with both Michie and Turing, too took part in the BP discussions of ‘chess-playing’ machines. In fact, Jack Copeland notes that Good recalls Turing speaking about the concept of mechanizing chess in 1941, before Michie’s time at BP. Like Michie, Good was interested in both chess and ‘learning’ machines. He was an excellent chess player apart from being a brilliant mathematician. Good was aware of the challenge that Michie and Wylie made to Turing and Champerowne’s chess-machine, and suggested a way in which MACHIAVELLI could be beaten in a letter he wrote to Turing in 1948:

I visited Oxford last week-end. Donald showed me a 'chess machine' invented by Shaun and himself. It suffers from the very serious disadvantage that it does not analyse more than one move ahead. I am convinced that such a machine could play a very poor game, however accurately it scored the position with respect to matter and space. In fact it could easily be beaten by playing 'psychologically', i.e. by taking into account the main weakness of the machine. This could be done by deliberately complicating the position and entering into combinations.

The correspondences show how the discussions on mechanizing chess from the BP years, evolved during the post-war period. However, these were not just the result of Michie and Turing’s acquaintance with chess players during their time at BP.

They were also shaped very much by the kind of codebreaking work they were involved in.

Colossus [c 1944]. The first-of-its-kind digital codebreaking was used for ‘guided’ searches to break German messages encoded using the Lorenz teleprinter cipher. (c) Crown copyright

 Both Michie’s and Turing’s chess-playing machines involved taking the opponent’s move as input and then creating an output move in response. Both of the algorithms involved ‘searching’ for a good move after considering various possible single moves by ranking them based on variables such as the safety of the piece, and the value of opponent’s pieces that could be captured. This was a guided search for ‘moves’ or ‘solutions’ using an evaluating – or ‘heuristic’ – function to narrow the number of possibilities rather than completing an exhaustive general search through all possible moves. It was conceptually very similar to the machine-aided searches that codebreakers at BP had to conduct to break the German ciphers. Michie’s job, being a key member of Newmanry team that broke the Lorenz cipher using the digital codebreaking machine Colossus, was to come up with solutions to narrow down the daily codebreaking searches of the machines. Such a problem-solving technique that uses a heuristic function to estimate hopeful paths to a solution would eventually be termed as "heuristic search". Michie himself would go on to make a significant contribution to developing and popularising this computational technique during the 1960s with his graph traverser program co-developed with J. E. Doran.

Going through Michie’s archives, we can find that research into chess-machines remained relevant to his entire machine intelligence and AI research career, even during the period that he worked as a geneticist. For instance, during the 1950s he had played MACHIAVELLI against biologist John Maynard Smith’s SOMA (Smith’s One-Move Analyser) in matchup refereed by Maynard Smith’s eldest son. A detailed description of how the machines performed was given in the 1961 New Scientist that Michie wrote together with Maynard Smith: ‘Machines that play games’.

Importance of chess to AI and machine learning

Chess became central to Michie’s research in the 1970s, when his research program was curtailed by the University of Edinburgh in the aftermath of the Lighthill Report that drastically defunded AI research in the UK. Chess endgames research was something Michie could work on with the limited funding he had post-Lighthill. However, for Michie, chess was not just a convenient and playful way to engage his interest in machine intelligence, as he clearly puts it in a 1980 draft titled ‘A representation for pattern knowledge in chess end-games’. In it, he addresses the question of whether those involved in computer chess are just ‘fooling around with the taxpayer’s money,’ and argues that ‘no other equally apposite material is readily available for investigating certain scientific issues of importance.’

This is also a point he emphasised in his 1966 paper, ‘Game-playing and game-learning automata’, referring to Turing’s interest in machines that could play games,:

It is sometimes thought that Turing’s interest in mechanised game playing was the spare time frivolity of a man who reserved his serious thoughts for worthier topics. That was not the case. He had the conviction that the development of high-speed digital computing equipment would make possible the mechanisation of human thought processes of every kind, and that games constituted an ideal model system in which studies of machine intelligence could first be developed.

Chess, for Michie, was the “fruit-fly” of AI research that was perfect for “studying the representation and measurement of knowledge in machines.’’ In another 1980 article ‘Chess with Computers’, he describes in detail why the strategic game is ideal for AI research and its “chief advantages”:

…chess constitutes a well-defined and formalized domain; it challenges the highest levels of intellectual capacity over a wide range of cognitive functions logical concept-formation, calculation, rote-learning, analogical thinking, deductive and inductive reasoning, and so forth; a detailed corpus of chess knowledge has accumulated over centuries in chess instructional works and commentaries; a generally accepted numerical scale for performance is available in the USCF rating system; and finally, the game can readily be decomposed into sub-games which can be subjected to intensive separate analysis.

Michie’s focus on chess-endgames research would eventually contribute to the development of the groundbreaking Iterative Dichotomiser 3 (ID3) learning algorithm for generating decision trees by J. Ross Quinlan, who was one of the many brilliant researchers Michie mentored during his career. In Quinlan’s 1986 paper, he acknowledges Michie’s role in the endeavour:

ID3 (Quinlan, 1979, 1983a) is one of a series of programs developed from CLS in response to a challenging induction task posed by Donald Michie, viz. to decide from pattern-based features alone whether a particular chess position in the King-Rook vs King-Knight endgame is lost for the Knight's side in a fixed number of ply.

This shows how chess shaped the fields of AI and machine learning; chess endgames research also played a key role in the projects of many of Michie’s PhD students from the 1970s and 80s, including Stephen Muggleton, Alan Shapiro and Tim Niblett’s pioneering work on induction methods in machine learning.

Nevertheless, the relationship between chess and AI research remains severely underexplored in the history of AI.

Posted by Aswin Valsala Narayanan

Further reading:

Donald Michie, “Alan Turing’s Mind Machines,” February 8, 2008, https://doi.org/10.7551/mitpress/7626.003.0005.

Herbert A. Simon and Allen Newell, “Heuristic Problem Solving: The Next Advance in Operations Research,” Operations Research 6, no. 1 (1958): 1–10.

Donald Michie, “Game-Playing and Game-Learning Automata,” in Advances in Programming and Non-Numerical Computation (Elsevier, 1966), 183–200.

Donald Michie, “Chess with Computers,” Interdisciplinary Science Reviews 5, no. 3 (January 1, 1980): 215–27, https://doi.org/10.1179/isr.1980.5.3.215.

J. R. Quinlan, “Induction of Decision Trees,” Machine Learning 1, no. 1 (March 1, 1986): 81–106, https://doi.org/10.1007/BF00116251.

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.

Aswin Valsala Narayanan is as PhD student at the University of Leeds and the British Library. He is on an WRoCAH Collaborative Doctoral Award researching the Donald Michie Archive, exploring Michie's work as an artificial intelligence researcher in post-war Britain.

The fiftieth anniversary of the foundation of the British Library is an opportunity to look back at the leading role the Library and its parent bodies played in introducing computerised information retrieval for science and medicine to the UK. Between 1965 and 1975 experiments in searching databases of medical research were carried out in partnership with the US National Library of Medicine (NLM)  together with computer scientists and medical users in the UK. Following the success of these experiments the Library launched BLAISE (British Library Automated Information Service)  as a national public service in 1977.

The NLM began publishing Index Medicus, an index of medical journal articles, in 1879. In 1960 printing was computerised and the machine readable data on tape became available for information retrieval. A publicly available US service, MEDLARS (Medical Literature Analysis and Retrieval System) opened in 1963 with MEDLINE  (MEDLARS online) going live in 1971. [1]

 In 1965 the NLM agreed with the National Lending Library for Science and Technology [2] to supply tapes in exchange for MEDLARS records of UK medical literature. With these tapes in hand the Office of Science and Technology Information [3]  funded Newcastle University to develop a retrieval package based on NLM’s IBM software to run on the university’s English Electric computer. Subsequent projects in 1973-74 tested the online environment and current awareness services with medical researchers and librarians in Leeds and Manchester over an online telephone link. [4]

The next step in service delivery was to establish online access to the NLM. University College London had set up a link to the US through ARPANET, the early version of the internet [5], and in 1973 British Library Research & Development [3] along with other public bodies, joined this network. This programme was historically significant as the first international communication over the internet. Project STEIN (Short Term Experimental Information Network) involved sixteen centres (e.g. the Royal Post-Graduate Medical School) each with its own terminal and trained intermediary.  The number of users (362) and searches (1217) was substantial and the study confirmed the need for intermediaries who were experienced in using the system and formulating searches. The clinicians and researchers who accompanied each session evaluated the results and gave feedback. Despite difficulties with telecoms, satisfaction was high as searches delivered articles that were new together with articles that were familiar to the users, thus increasing their confidence in the search. [6]

These encouraging results led the Library in 1977 to launch BLAISE, a fully supported public service providing Medline and databases for toxicology and cancer. Tapes were delivered monthly from Washington by diplomatic bag to a computer bureau in Harlow to run on an IBM 370 machine with NLM’s ELHILL retrieval software. Mounting tested software on an established bureau service meant that BLAISE went live within a year. Users benefited from the dedicated BLAISE PSS (Packet Switched Service) network and a support team that provided training, documentation and a help desk, alongside document supply from the British Library Lending Division at Boston Spa.[7] At first researchers and clinicians used Medline for checking references or keeping up to date but it has since become an essential tool for the evidence based medicine community to generate systematic reviews and contribute to the Cochrane Library.[8] From 1977 the Library was the sole provider of NLM databases in the UK but in a political decision in 1982 NLM, as a federal agency, was required to release its products to US online providers. With the ensuing competition BLAISE was no longer able to support a UK based service and it was relaunched as BLAISE-LINK, a UK portal for online access to NLM. Within a few years customers moved over to commercial online hosts and BLAISE-LINK closed. 

Today, the Library continues online healthcare with the publication of AMED (Allied and Complementary Medicine Database). This database supplements the coverage of Medline in areas such as alternative medicine, palliative care and rehabilitation. [9]

We have come a long way in fifty years.  In 1973 searching involved expensive telecoms and computer access, clumsy equipment  (who now remembers audio-acoustic couplers?) minimal records, complex Boolean search strings and the need for skilled medical librarians to navigate all these obstacles. Now, there is free access to the internet and PubMed, open access full text and sophisticated relevance searching empowering every user. Information has exploded:  in 1976, Medline and its associated files had 3.5 million records, by 2022, PubMed had 35 million. [10] 

References [BL shelfmark]

All URLs accessed on 7 July 2023.

[1] MEDLINE History. https://www.nlm.nih.gov/medline/medline_history.html

[2] The National Lending Library for Science and Technology (NLLST) was the predecessor of the British Library Lending Division and later, the Document Supply Centre. The service is currently available as British Library On Demand.

Barr, D. P. The National Lending Library for Science and Technology. Postgraduate Medical Journal. 42.493 (1966): 695. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2466097/pdf/postmedj00407-0003.pdf

[3] The Office of Science and Technology Information (OSTI) was the predecessor of British Library Research & Development which promoted and funded R&D by the UK library and information community until its merger with the Library and Information Commission in 1999.

Baxter, P. "The role of the British Library R&D department in supporting library and information research in the United Kingdom." Journal of the American Society for Information Science 36.4 (1985): 276. https://www.proquest.com/openview/77a69cbd42dd0412f39b217892f95ac2/1?pq-origsite=gscholar&cbl=1818555 

[4] Barraclough, E. Information Retrieval, its origins in Newcastle. http://history.cs.ncl.ac.uk/anniversaries/40th/webbook/infoRetrieval/index.html

Harley, A. J., and Elizabeth D. Barraclough. MEDLARS information retrieval in Britain. Postgraduate medical journal 42.484 (1966): 69. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2465839/pdf/postmedj00398-0003.pdf

[5] Kirstein, P. T. "Early experiences with the Arpanet and Internet in the United Kingdom." IEEE Annals of the History of Computing 21.1 (1999): 38-44. https://citeseerx.ist.psu.edu/doc/10.1.1.112.8527

Computer History – Internet history of the 1970s.  https://www.computerhistory.org/internethistory/1970s/

[6]  Holmes, P. A description of the British Library’s short-term experimental information network project. pp 231-237 - 1st International On-line Information Meeting, London 13-15 December 1977 / organised by On-line Review, the international journal of on-line information systems. (1977). Oxford ; New York: Learned Information. [available in the British Library at shelfmark 2719.x.4085 ]

Holmes, P. (1978). On-line information retrieval: An introduction and guide to the British Library's short-term experimental information network project / P.L. Holmes. Vol.2, Experimental use of medical information services. (Research and development reports (British Library) ; no.5397). London: British Library Research and Development Department. [available in the British Library at shelfmark 2113.560000F BLRDR 5397 ]

Trials were also made with other scientific and engineering databases on the Lockheed Dialog system.

(7) Holmes, P. L. The British Library Automated Information Service (BLAISE). Online Review 3.3 (1979): 265-274. https://www.emerald.com/insight/content/doi/10.1108/eb024003/full/html       

BLAISE also provided bibliographic databases for the British National Bibliography and the Library of Congress, finally closing in 2002.

[8] McKibbon, K. A. Evidence-based practice. Bulletin of the medical library association 86.3 (1998): 396. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC226388/pdf/mlab00092-0108.pdf

Cochrane Library. https://www.cochranelibrary.com/about/about-cochrane-reviews

[9] Allied and Complementary Medicine Database (AMED) https://www.ebsco.com/products/research-databases/allied-and-complementary-medicine-database-amed

[10] Miles, W. (1982). A history of the National Library of Medicine : The nation's treasury of medical knowledge. (NIH publication ; no.82-1904). Bethesda, Md.: U.S. Department of Health and Human Services, Public Health Service, National Institutes of Health, National Library of Medicine. [p.386 -3.5 m records, 1976] https://collections.nlm.nih.gov/bookviewer?PID=nlm:nlmuid-8218545-bk

PubMed Milestone - 35 Millionth Journal Citation Added. https://www.nlm.nih.gov/pubs/techbull/nd22/brief/nd22_pubmed_milestone.html

Further reading

Bourne, C., & Hahn, Trudi Bellardo. (2003). A history of online information services, 1963-1976, Cambridge, Mass. ; London: MIT. [Available in the British Library on open shelf: Humanities 2 Reading Room HUR 025.04]

Written by Richard Wakeford (Science Reference Specialist, Retired). Richard was a member of the BLAISE support team, 1981-1984.