A Bit More Detail

Assorted Personal Notations, Essays, and Other Jottings

[LINK] The Yorkshire Ranter on how computer models nuclear started a nuclear war in 1983

In his post “That time I was nearly burned alive by a machine-learning model and didn’t even notice for 33 years”, Yorkshire Ranter Alex Harrowell writes about how subtle issue with Soviet computer models of Western behaviour nearly started a nuclear war in the early 1980s. Starting principles, it seems, must be carefully examined.

We already knew about Operation RYAN, the Yuri Andropov-inspired maximum effort search for intelligence offering strategic warning of a putative Western preventive war against the Soviet Union, and that it intersected dangerously with the war scare of 1983. We also knew that part of it was something to do with an effort to assess the intelligence take using some sort of computer system, but not in any detail. A lot more documents have just been declassified, and it turns out that the computer element was not just a detail, but absolutely central to RYAN.

At the end of the 1970s the USSR was at the zenith of its power, but the KGB leadership especially were anxious about the state of the economy and about the so-called scientific-technological revolution, the equivalent of the Revolution in Military Affairs concept in the US. As a result, they feared that once the US regained a substantial advantage it would attack. The answer was to develop an automated system to predict when this might happen and what the key indicators were.

Model the whole problem as a system of interconnected linear programming problems. They said. Load up the data. They said. Comrades, let’s optimise. They said.

In all, the RYAN model used some 40,000 data points, most of which were collected by greatly increased KGB and Joint GRU field activity. It generated a numerical score between 0 and 100. Higher was better – above 70 peace was probable, whereas below 60 it was time to worry. The problem was the weighting applied to each of those parameters. Clearly, they had to train the model against some existing data set, and the one they chose was Nazi Germany in the run-up to Operation BARBAROSSA.

Who needs theory? They said. We’ve got the data. They said. A simple matter of programming. They said.

Written by Randy McDonald

December 7, 2015 at 11:59 pm