More Babbage in "Jeopardy!"

BABBAGE had $5,000 to his credit in the final moments of his third match of the game show "Jeopardy!" that aired a few weeks ago. The category was 19th-century female authors. Trailing the leader by $3,600 with a couple of minutes left in the game, he was faced with a "Daily Double" wager that allowed him to bet from $5 to the full $5,000 in his kitty. A fraction of a second later he bet the whole lot. Would IBM's Watson supercomputer, which in 2011 defeated the programme's two all-time best human contestants, have done the same?

Almost certainly, according to Gerald Tesauro, a researcher at IBM, and four colleagues. Earlier this year they published a paper in the firm's research journal detailing the approach to wagering that helped Watson beat Ken Jennings (the longest-playing contestant) and Brad Rutter (its highest-grossing winner) in a series of televised matches.

The emphasis at the time was on the extraordinary breakthroughs in natural-language processing that allowed real-time parsing of the clues and (mostly correct) responses, which the programme requires to be phrased in the form of questions. But the authors, part of a team of two dozen scientists who worked on the "Jeopardy!" project over four years, explain that tactical game-play choices based on probabilities of the answers being correct also played a big part. Watson, like any player, had to pick which hidden clue at a given dollar value to select, predict where the doubling clues might lie and calculate amounts to wager during Daily Doubles. Dr Tesauro says the computer had a big advantage over human contestants, even the best of whom focus on factual knowledge and largely ignore the game's probabilistic facets (which Babbage can confirm).

For a start, Dr Tesauro's group put a number on what viewers and players know: that Daily Doubles are more likely to hide behind certain squares on the board than others. In the second round, Double Jeopardy, the show's producers never place them both in the same category. In both rounds, producers located them typically in one of the highest three dollar values. In either round, one often appears in the first category column. From this, a prediction model helped Watson pick Daily Doubles when it was most advantageous.

Researchers also modelled typical "Jeopardy!" players to simulate games against Watson by mining the fan-built J-Archive repository of decades' worth of clues and responses. But rather than look at the content of the trivia, the researchers looked at the order of selection, wagers and accuracy. The team created models of average players, champions (semifinalists from an annual tournament of the year's best players) and grand champions (those who had won the most games in the current show's run).

Unsurprisingly, champions and grand champions ring in for regular clues much more often than average contestants, and are more accurate. But the difference is much higher in Daily Double and the last round called Final Jeopardy. For instance, a Ken Jennings would be correct 82% of the time with a Daily Double wager, while the average contestant had the right response in only 64% of cases. "It gave me a new way of appreciating just how good Ken Jennings is", says Dr Tesauro.

Dr Tesauro says the IBM system wound up playing a very different game than its human counterparts. Something similar has been seen in more traditional games, such as chess, backgammon and bridge, once algorithms became robust enough to challenge master-class players. For instance, when Watson wrested control of category selection it ventured all across the board rather than work its way through a single category (which the show's producers encourage but cannot enforce).

While typical contestants wager even amounts for Daily Doubles, like $1,000, and rarely above $5,000, Watson bet sums like $1,246 or $6,435. Dr Tesauro explains that the software's clue analysis component would score how confident Watson was in an accurate answer which would be used to shape the size of a wager. The wagers factored in the rounds' limited time, too. The paper also reveals that picking the Daily Double clues at the right moment can tip the scales. In a simulation of 200,000 games against grand champions, Watson could win 61% of the time with top-to-bottom selection, and 68% of the time by using multiple methods to pinpoint the wagering rectangles.

This is all fun and games but, as IBM said at the time, to a point. Dr Tesauro says that the ability to model—and beat—human decision-making processes has applications in any field where different decisions lead to a multitude of possible outcomes. In a paper for an upcoming conference, Dr Tesauro and his colleagues extend their "Jeopardy!" thinking into health care, where varying incentives for doctor compensation and disparate treatment possibilities were weighed to find the optimum solution. Such software would potentially let policymakers, regulators and hospital managers model scenarios without having to test them out on real patients.

As for "Jeopardy!", Dr Tesauro expects (and hopes) that Watson's approach will eventually colour human players' preparation and gameplay. He suggests that alongside cramming knowledge, potential contestants should mug up on game theory. (He also acknowledges that "Prisoner of Trebekistan", a book written by Bob Harris about his multiple campaigns, dovetails neatly with their more exhaustive data analysis.)

So, how did Babbage do? Faced with recalling the pen name of Amantine Lucile Aurore Dupin, his addled brain came up with "George Saunders" (a contemporary author), obviously incorrect. The remnants of his cognitive function translated this to "George Sands". Sadly, the answer was "Sand". Your correspondent lost the lot (though he regained $2,000 before Final Jeopardy; the right answer there was not enough to win). Dr Tesauro ran the scenario through algorithms that simulate only human players, and says that with the degree of confidence Babbage had in the category risking it all was a good strategy. Some solace.