Hacker News new | past | comments | ask | show | jobs | submit login

Keep in mind that it's using time travel to do those moves. In the real world time travel is not currently available.



What exactly does "time travel" mean in this context? That the simulation makes some moves, sees what happened, and then has the ability to undo its moves and go back to its old state, and then try new moves?


Yes. It's basically a brute-force Groundhog Day approach to finding the best move.


Another movie analogy would be Next. Cris Johnson can see two minutes into the future and redo anything that has a bad outcome.

http://en.wikipedia.org/wiki/Next_(2007_film)


Yes, exactly.


When a baseball player swings a bat, in anticipation of the ball arriving just in time to get knocked out of the park, do you consider that time travel?

Is the baseball player's anticipation of the future really all that different from what this program is doing?


Yes, it's incredibly different! A huge problem in AI is internally modelling the outside world. This completely sidesteps that.


Maybe, but that doesn't really have anything to do with the "time travel" question.


Of course it does. Models are useful because they have predictive power. If you have time travel, you don't need to predict anything.


I think the difference is that the computer can try the input and record the result and then try again. The baseball player can't have a redo of that pitch if they miss or if it's a foul.

I don't really see how it's time travel but that's what he called it.


the program fails repeatedly and then loads a previous state to try a different move, what you see is the result of multiple trials. it can't reliably get it right on the first try.


Neither can a baseball player. He's been practicing since age 3 to hit that ball.

Is a decade or two of baseball practice really all that different from a computer playing and re-playing the game over and over looking for optimal strategy?


It is if it can't learn from those mistakes.


Simulation and models often are, though.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: