Google DeepMind gaming AI just found another way to make code faster

Google DeepMind gaming AI just found another way to make | itkovian

DeepMind compares AlphaDev’s discovery to one of AlphaGo’s strange but winning moves in his Go match against grandmaster Lee Sedol in 2016. “All the experts looked at this move and said, ‘This is not the right thing to do. . This is a bad move,’” Mankowitz says. « But it was actually the right move, and AlphaGo ended up not only winning the game, but also influencing the strategies that professional Go players started using. »

Sanders is impressed, but doesn’t think the results should be overstated. “I agree that machine learning techniques are increasingly a game changer in programming and everyone expects that AIs will soon be able to come up with new and better algorithms,” he says. « But we’re not quite there yet. »

First, Sanders points out that AlphaDev uses only a subset of the instructions available in assembly. Many existing sorting algorithms use instructions that AlphaDev hasn’t tried, he says. This makes it more difficult to compare AlphaDev to top rival approaches.

It is true that AlphaDev has its limitations. The longest algorithm it produced was 130 statements long, for sorting a list of up to five items. At each step, AlphaDev chose from 297 possible assembly instructions (among many others). « Over 297 instructions and 130+ instruction long assembly games, learning just got slow, » says Mankowitz.

This is because even with 297 instructions (or game moves), the number of possible algorithms that AlphaDev could build is greater than the possible number of games in chess (10120) and the number of atoms in the universe (about 1080).

For longer algorithms, the team plans to adapt AlphaDev to work with C++ statements instead of assembly. With less granular control AlphaDev might lose some shortcuts, but the approach would be applicable to a wider range of algorithms.

Sanders would also like to see a more comprehensive comparison of the best man-made approaches, especially for longer algorithms. DeepMind says it’s part of his plan. Mankowitz wants to combine AlphaDev with the best man-made methods, making AI build on human intuition rather than starting from scratch.

After all, there may be more accelerations to be found. « To do this, it takes a human being a lot of experience and an enormous amount of hours, maybe days, maybe weeks, to go through these programs and identify improvements, » says Mankowitz. « Consequently, it hasn’t been attempted before. »

Hi, I’m Samuel