What Does Prediction And Decision Making Mean?

What Does Prediction And Decision Making Mean?


It is 50 years to yestereday, since the brilliant but unpredictable Bobby Fischer defeated Boris Spassky to win the “Match of the Century”, on Sep 1 1972.

The match was echoed in the musical “Chess”, and reflected in the Netflix series “The Queen’s Gambit”. Sadly, both players sunk to tragic lows years later (in Spassky’s case: the ignominy of playing myself in a noisy Sydney shopping center).

The most famous game from the most famous match of all time was Game 3.

The most famous move from the most famous game of the most famous match of all time was 11 …Nh5 by Fischer. You can buy Nh5 t-shirts. I have one. (I also stopped by the obscure Bobby Fischer center in Iceland a few months ago but alas, no one was home to help me add to my swag).

IBM, Deep Blue & Kasparov : A Chess Story

Chess provides a rather interesting demonstration of value functions, the critical link between prediction and decision making. In theory if an oracle can give you all the conditional positional evaluations, you can play chess.

Oh, but where do you find one?

Well, in theory, the world will one day be full of oracles … although they will work much better for fast moving manufacturing processes, IoT, financial markets and the like, than they do for chess, admittedly.

It is a nuanced issue though. For instance, if you put the most famous game of all time into a strong open-source computer (an oracle substitute for the sake of argument) the sequence of valuations is not in accordance with the historical narrative.

Emotionally, we want to believe that 11 …Nh5 was situationally brilliant but without exogenous data (measuring the beads of sweat on Spassky’s brow, etc) it is difficult to get computers to agree. At some level this is a banal observation, the difference between expectation and minimax.

a close up of text on a white background

But more generally, how do you make decisions if you partially trust the value function? Perhaps we should borrow some tricks from temporal difference learning, or elsewhere.

As far as I know, there isn’t a lot of discussion about the collision between crowd-sourcing and reinforcement learning. But there is some in my forthcoming book (https://lnkd.in/eRWBBrge). And I hope some of you care to debug my thoughts on the topic!

Written by Peter Cotton

What Does Prediction And Decision Making Mean?