Masayuki Mochizuki

And

I'll just keep coming back and giving you my money like an ATM.

That's an interesting conundrum.

I'd love you to explain when the neural networks came along and you could actually identify what were good moves and what were bad moves and you could acquire what's called an error rate or performance rating.

What was that like for you?

Before the neural nets, nobody really knew what the best plays were.

That's Mark Olson again.

Backgammon Galaxy is built atop a neural network that plays the game at a superhuman level.

I asked him how this technology came about.

In the 90s, AI algorithms with reinforcement learning was invented.

A guy called Gerald Tesoro invented a program called TDGammon.

This was the first neural net to play backgammon really well, stronger than even the best players in the world.

Where did this come in the development of AI play?

This was basically around the time when Deep Blue were playing against Kasparov.

Gerald Tesaro used backgammon as a...

case study for his new AI algorithm, but Gammon was such a perfect environment for that because it's a multi-dimensional game where all these dimensions interact with each other because there's so many trade-offs and the variables are entangled, yet it's still a small enough universe that you can just reproduce data

as much as you want.

You can simulate the games.

And it worked exceptionally well.

And when you say it worked, that means that in any given moment in any game, there is an optimal move, essentially, correct?