Post by talinsalway on Feb 3, 2014 17:11:56 GMT -8
I've had a thought that's been kicking around -
A lot of the work I've seen so far, and the last few weeks of the coursera course, seemed to be more focused on improving performance of state evaluations, as opposed to more intelligently exploring states. For example, propnets trade some logical power for vastly faster state evaluation. (That said, propnets can also be an easier way to find certain factorings).
Now, it is true that humans and computers have different strengths when it comes to problem solving or game play - humans currently are better at finding patterns and filtering out irrelevant actions, even when those actions have some impact on the game and couldn't be factored out, while computers can look at thousands more states per second than a human can. That said, I think it's worth trying to develop these capabilities for computer players.
I'd like to propose an alternative to the GGP format, where instead of each turn being time-limited, each turn is limited by the number of states you're allowed to explore (probably including a reasonable time limit to do so in).
The advantages are - developers spend more time optimizing for factorings and logical analysis of game states, as opposed to efficiently evaluating them for the basic goal & legal move set. There's less of an advantage to using a low-level performance oriented language, so more experimentation will be done in higher-level, less-efficient languages.
There's many disadvantages, primarily - it might be hard to define what exactly constitutes a state evaluation, and enforcement of such a rule would primarily be on the honor system (except in structured tournaments where source code is readily available for analysis). Also, players would have to have separate versions, or runtime configuration, to know whether they should be limited by time or by evaluations, and there would be changes to existing GGP infrastructure.
A lot of the work I've seen so far, and the last few weeks of the coursera course, seemed to be more focused on improving performance of state evaluations, as opposed to more intelligently exploring states. For example, propnets trade some logical power for vastly faster state evaluation. (That said, propnets can also be an easier way to find certain factorings).
Now, it is true that humans and computers have different strengths when it comes to problem solving or game play - humans currently are better at finding patterns and filtering out irrelevant actions, even when those actions have some impact on the game and couldn't be factored out, while computers can look at thousands more states per second than a human can. That said, I think it's worth trying to develop these capabilities for computer players.
I'd like to propose an alternative to the GGP format, where instead of each turn being time-limited, each turn is limited by the number of states you're allowed to explore (probably including a reasonable time limit to do so in).
The advantages are - developers spend more time optimizing for factorings and logical analysis of game states, as opposed to efficiently evaluating them for the basic goal & legal move set. There's less of an advantage to using a low-level performance oriented language, so more experimentation will be done in higher-level, less-efficient languages.
There's many disadvantages, primarily - it might be hard to define what exactly constitutes a state evaluation, and enforcement of such a rule would primarily be on the honor system (except in structured tournaments where source code is readily available for analysis). Also, players would have to have separate versions, or runtime configuration, to know whether they should be limited by time or by evaluations, and there would be changes to existing GGP infrastructure.