WHAT’S NEXT FOR ALPHAZERO?
DeepMind’s AlphaZero has already overcome rival AI engines and inspired grandmasters to change their style of play. So what else is left for AlphaZero to achieve? After all, it’s not as if any human grandmasters will be stupid enough to take it on. As Nigel Short remarks, even “your mobile phone will just thrash the world champion” these days.
DeepMind’s staff research engineer, Nenad Tomasev, says the project is now focused on giving something back to the chess community. “We are no longer actively improving AlphaZero on chess, to make it stronger,” he said. “Instead, we see this as an opportunity to do research on identifying ways in which AIs can best complement and enrich human skill, as this is likely to be important in various domains of future AI applications.”
That doesn’t mean they’ve given up on chess entirely. The team are working with ex-world champion Vladimir Kramnik to explore various projects, such as prototyping alternative rule sets for chess. As Nigel Short mentioned, the team has explored including stalemate as a victory as well as more fundamental changes. These include eight other rule changes, such as no-castling, pawns being able to move backwards as well as forwards and the option to capture one’s own pieces. The paper exploring the new variants can be found at pcpro.link/330zero1.
A second project is, says Tomasev, “about understanding how AlphaZero learns and encodes chess knowledge, in order to develop ways of accessing pieces of it that could help inform the human understanding of the game”. The paper ( pcpro.link/330zero2) shows that there is limited overlap between the way AlphaZero learned the game and the way humans do, which could have big implications for the way the game and other concepts are taught by AI.