Assessing the Military Implications of AlphaGo: A Year On

The past week (May 22, 2017) marks the final appearance of AlphaGo – as Deep Mind retires it from professional play.[1] In little over a year, AlphaGo has played three sets of games against the top professional players (5 against Lee Sedol, 60 against professionals online, and 3 against Ke Jie) – in roughly 6 month intervals – with an astounding record of 67 to 1.

The progress of AlphaGo has been staggering – with Michael Redmond[2] noting that in each set of games played, AlphaGo advances so much that it may as well be a completely different player. In terms of quality of play, the style of AlphaGo has been not only innovative, but also hugely influential. It has invented entirely new standard play sequences (joseki), has revived some of the old sequences that were believed to be subpar; it has shifted the style of play towards influence – rather than territory, and provided the Go community with a whole new take on shifting positional play. Merely a year later, Go professionals commonly play AlphaGo moves.

Part of this flowering of innovation is, of course, due to the kind of processing power and analysis that AlphaGo has brought to bear on the game – with unprecedented ability to integrate, data-mine, and compare the game records of professional play over the last 400 years or so – not to mention the tens of thousands of on-line game records available. Further, the AlphaGo that played against Ke Jie is a much more streamlined version – using only about 10% of the computing capacity of the iteration that played against Lee Sedol – despite its ever-advancing gameplay.

Perhaps the best indicator of AlphaGo’s progress over the 14 months is the fact that very few – if any – professionals believed that Ke Jie (the top ranked player in China, and in the top 3 players in the world) could actually beat the machine. By comparison, in March 2016, there was serious skepticism among the Go professionals that AlphaGo could take on Lee Sedol.

The military implications of this rate of development are equally staggering. The possibility of militarizing AlphaGo into AlphaWar have already been discussed elsewhere.[3] However, the rate of progress, assimilation of data, and the self-learning capacities make a reevaluation important for a number of reasons.

First, since AlphaGo’s debut, a number of other AI systems have made their appearance on line, and are now regularly beating professional players. To a certain extent, this was to be expected; when the “impossible” is achieved once, replication becomes common. In practical terms, this means that the militarization of AlphaGo is no longer dependent on Deep Mind’s willingness to allow their technology to be militarized; the DOD (and other organizations) can just as easily obtain the rights for another piece of software, or build their own.

Second, the number of wars we’re engaged in – and are likely to become engaged in – is growing: Afghanistan, Iraq, Syria, Libya, Somalia, Yemen, ISIS, and of course the increasing possibility of N. Korea and Iran. Within that context, the use of drones is also on the rise, where the drones used are both military, as well as off-the shelf variety. ISIS, for example, has been making increased use of off-the-shelf drones for both reconnaissance and payload delivery.[4] This rise of drone use means that drones are likely to play a more significant part of all future warfare, both in terms of the US, but also in terms of other states and entities – including Russia, China, ISIS, Israel, and Iran.

As an additional factor of the increased global violence, it is also worth noting that, for the first time in a long while, the major powers are coming into relatively direct conflict. China’s expansion into the South China Sea is notable, as is the US-Russia involvement in Syria, where the escalation of violence has included the April 6 US attack on the Assad regime with 59 Tomahawk missiles, and the May 18 strike on a Syrian military convoy. Both attacks were seen as a direct provocation by Russia, a close military ally of Syria.

Third, the data collection grows daily, and the interconnectedness that allows for data-mining does so as well. This provides a vast store of learning materials for AlphaWar. Additionally, the documentation of various rebel and government forces, tactics, and capacities is growing as well, providing further datasets to be studied – including their capacity for development and use of drones. The kind of in-depth analysis available to AlphaGo has already moved it by leaps and bounds beyond human capacity. The ability to simultaneously inspect the entire array of known data and match it against the present situation, to find hidden relations, to test thousands of options in seconds, and to act without emotional irrationality, have already demonstrated themselves to be functional, actionable, and winning strategies in outsmarting the most capable humans.

It may be telling that the only game AlphaGo has lost professionally required something of a Deux ex Machina (pun intended), when it lost game 4 against Lee Sedol. Lee Sedol’s move 78 is aptly known as “Hand of God” (Kame no Itte), and is the kind of move that top players may manage to play once in their entire career. That is, the only loss suffered by AlphaGo required the absolute pinnacle of human ability – an insight so profound that it is essentially attributed to a superhuman power.

Fourth, the development of drones themselves has been exponential. New models of commercial drones – to say nothing of potential military developments – are capable of self-organization, fluid formation flight, amazing acceleration (0-80 in under a second) with a carrying capacity of 15 lbs., etc.[5] Combined with basic features like facial recognition, it may soon be possible to send out a flock of mini-drones to find and identify specific individuals as targets for precision strikes.

The potential for AlphaWar utility on the battlefield is undeniable, especially once both sides in a conflict start using drones on a larger scale. Drone control is an easy factor to manage; the difficult part is making use of the yottabytes of data and metadata, and combining the variety of variables into a single coherent process by which military operations are structured. This process of ordering chaos is one that has been the core of military strategy development. However, as wars have moved away from the traditional “armies facing each other on a field” model, the actionable analysis of data has become far more complex – and may be functionally beyond unassisted human capacities. From what we have seen of AlphaGo, this type of in-depth analysis, establishing of correlations too obscure for human recognition, ability to parse through lifetimes of data in months, and evaluation of realities and possibilities on the board are precisely the attributes this type of learning algorithm has mastered.

Given the utility and capacity of AlphaWar, it would be rather surprising if at least one major power (US/Russia/China) does not have a functional model of AlphaWar up and running by 2020. In fact, there’s little reason to assume that the race is not already on – with the first developer gaining a notable military edge (an issue that seems crucial, given the increasingly contested nature of several global strategic regions between the major powers).

The rise of the drones, of course, also opens up an entirely new area of warfare against the drones. Given the remote operation basis of drones, and the vast attack surface such connections present, we can expect that drone hacking and drone hijacking – hellfire missiles and all – will soon come to the fore. This is likely to create a whole new market for cyber security (CS) services, and will present a problem for the US specifically.[6]

From a technical and ethical standpoint, as well as the standpoint of Go, there are interesting times ahead. However, we should remember that, for the Chinese, “May you live in interesting times” has been a traditional curse.

 

[Originally published: https://www.academia.edu/33269676/Assessing_the_Military_Implications_of_AlphaGo_A_Year_On]

[1] While retiring AlphaGo, Deep Mind is set to release 50 games AlphaGo played against itself, as well as AlphaGo’s analysis of Go positions. https://deepmind.com/blog/alphagos-next-move/

[2] http://senseis.xmp.net/?MichaelRedmond

[3] Rahmanovic, Faruk and Connor Applegate. Military Implications of AlphaGo. https://www.academia.edu/32068311/Military_Implications_of_AlphaGo

[4] Warrick, Joby. “Use of Weaponized drones by ISIS spurs terrorism fears.” The Washington Post. February, 21, 2017. https://www.washingtonpost.com/world/national-security/use-of-weaponized-drones-by-isis-spurs-terrorism-fears/2017/02/21/9d83d51e-f382-11e6-8d72-263470bf0401_story.html?utm_term=.dd8abd9504c1

[5] Darcy, Kieran. “DRL’s next-gen Racer3 drone combines speed, performance.” April 7, 2017.  http://www.espn.com/espn/story/_/id/19100909/drone-racing-league-introduces-next-generation-racing-drone

[6] As has been commonly noted by the CS community, the US is already suffering a severe deficit of capable CS personnel. Part of the problem stems from the fact that the traditional foreign sources of security experts (Russia, China, India) have been growing, and are now able to retain more of their own experts – reducing the supply of such individuals to the US markets.
Rahmanovic, Faruk. “Dangers of a False Narrative.” Actionable Cyber Security. April 21, 2017. http://actionablecs.com/dangers-of-a-false-narrative/  And
Rahmanovic, Faruk. “Security is not a Product.” Actionable Cyber Security.  April, 21, 2017. http://actionablecs.com/security-is-not-a-product/

Leave a Reply

Your email address will not be published. Required fields are marked *