Computers are very clever beasties – at least most of the time. Sadly their matches against each other are deadly dull. The games virtually never have strong strategic lines of thinking – which is the main thing that makes chess absorbing (for me anyway – a battle is waged: a story is told). Still, for my sins, I occasionally check out their games which are usually very long and involve some small tactical edge emerging and then gradually being converted into a win – or just winding down into a draw.
Chessbomb is a great site where you can watch chess tournaments from all round the world, move by move as the games unfold as well as a powerful computer analysis of the positions as they unfold. The computer is usually a very accurate judge of how things are going. Anyway, you can also watch computers grinding away against each other. One such tournament, underway now, is TCEC S1 Division 2. What does all that mean? I have no idea, but it’s a computers’ comp.
While working away on other things last night I kept the occasional eye on the game in round 9 between Protector 1.36.387 and Junior 12.0. 1 And something very strange unfolded. Protector was smart enough to get itself an entirely won game against Junior which is a seriously good program. When I checked out the game this morning Protector had flipped its lid. It didn’t seem to know that to endgames you should try to hang onto the pawns you have (they’re don’t make them any more after the start of the game.
No human beings were hurt in the hosting of the game (I presume) but some programmer’s bonus may have been cut. Seriously, the thing seems to have blown some gasket. I could have won the endgame against Junior!
- I’ve just figured out how to provide a direct link. NG[↩]
Maybe the name “Protector” reflects an algorithm that concentrates on preserving higher value pieces, and content to put raw prawns on the fire.
Slight errors in data configuration and weightings can lead to weird behaviour. I remember an early version of a medical history taker I wrote, a bank of thousands of questions with the strategy designed to ask the most pertinent multi-choice questions based on previous responses, aiming at a mere 15 minutes of effort for the patient (simple example – if female, you’d get asked about pregnancies) – deterministic, but difficult to predict the next question. It had a set of answers designed for early practice in the session, “bananas, pineapples, oranges”, but after getting confirmation that a person had been out of victoria at some stage, asked “any history of tropical diseases?” and then offered the list of tropical fruit as possible responses.
Even without frank bugs, complex goal-seeking programs can emulate not merely human intelligence, but human lapses of judgement.
More specifically for chess programs, the GNU engine has had a “coffee shop” mode… playing usually at a standard above it’s nominal level, with occasional lapses to two levels below, as if distracted by cake or a full bladder (it pushed you, but let you take advantage of those occasional opportunities if you were preepared to change tactics suddenly). Did you ever play this engine and get “brainos” that were particularly daft?
Well, looking at the leaderboard Protector is definitely the weakest of the computers, but then it’s got five points, and if most quite good players were playing in this tourney I think they’d have scored zero – with their opponents just grinding away picking up small inaccuracies.
Further, on the occasions that it gave away pawns, it didn’t have to. So it’s pretty inexplicable I think.
I’m still with the theory that Hal’s blown a gasket.