In other science news, a computer program defeated the Go champion.
In a sense, this is hardly a surprise: in 1910, Ernst Zermelo proved that all games of perfect information have a well-defined perfect “solution” that always leads to the same outcome. Depending on the specific rules, either one player has a perfect set of moves that cannot be defeated, or every set of moves has a perfect set of countermoves that leads to the game ending up in a permanent stalemate. The only complication that has prevented this perfect solution (or the nonexistence thereof) for Go or Chess from being discovered so far is the lack of computing power.
The lack of human computing power meant that good board game players have come to treat such complex games as if playing a game of imperfect information (like poker) where they have to make a series of judgments based on conditional probabilities: what are the top 10 likely possibilities conditional on some configuration of pieces on the board and what are the top 10 best countermoves to them, rather than try to force through calculations for all possible configurations, for example. This is really what “intuition” amounts to in nutshell: what we do to cheat having to go through all the calculations. This is easier when all the moves are defined explicitly and can be anticipated perfectly–if only you could do all the necessary calculations, at least in theory.
But all the intuition, at least within the rules of a board game, is still math: truncated, simplified math that skips over a lot of complexity with convenient simplifying assumptions–exactly the kind of stuff that engineers are very good at, in fact. This is how a lot of rules and formulas in engineering came about, after all. The same mindset can be applied to address how a computer “intuits” by optimizing computing power. Since a computer is still capable of vastly more complex calculations than a human, it can calculate top 1000 rather than top 10 conditional possibilities and calculate top 1000 countermoves rather than top 10 and so forth. Once “intuition,” with the aid of well-defined rules, becomes a fairly straightforward if still vastly complex mathematical problem,
The trouble comes from “well-defined rules.” Chess and Go are easy. Poker is harder, since its play is dominated by both rules and uncertainty–and how players can game the rules to mess with uncertainty, or the perception thereof. Still, poker is bound by the rules of the game and the extent to which players can game the rules can, in principle, anticipated. What about the broader game of human interactions, though? There is enough that is governed by reasonably well-defined, set rules. But with subtleties through which they can be exploited. I remain amazed by how Google translate operates, but I always wondered if that is the wrong way forward: as far as I know, it operates not by actually learning the words, but simply recognizing the patterns in their usage in vast pools of data. Since 99.9% of life is people going about life in predictable, “usual” fashion, just learning the patterns is good enough for 99.9% of life also. But sometimes, Sanders defeats Clinton in Michigan and those man-bites-dog events are significant, however rare and unusual they might be–and “rare” events are both more impactful and frequent than one might expect (e.g. the financial meltdown and the consequences which we are still living with).
This brings us back to the conceit of “data science” that I’m uncomfortable with, in its many applications. It encourages too much focus on the “average” (not literally, but the simple statistical quantifications of what is to be conditionally expected “normally.”). What about oddballs, unexpected outcomes, and other strange things? (Yes, I do obsess with conditional “variance” more than the conditional means.) When we depart from the realm of well-defined rules, conditional variances become more important. Will Google invent an algorithm for a computer to naturally mimic bad grammar artfully and tastefully? Perhaps so–people are conditionally more predictable than not, or so I hear. But finding patterns is easy when the rules are well-defined, it is finding anti-patterns that is more meaningful when stakes get bigger….