[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

"Easy" improvement to PR calculations

Posted By: Timothy Chow
Date: Tuesday, 16 April 2013, at 5:28 p.m.

In Response To: "Easy" improvement to PR calculations (Henrik Bukkjaer)

Henrik wrote:

In another thread (the one about error-rates and complexity in chess), you mentioned something about analyzing a heap of backgammon matches, in order to compensate the error-rate calculation (or something like that) for complexity. I assume categorizing positions into position types, and then analyzing which types typically were difficult to play for humans, etc.

I wonder how you would use that information (if gathered/derived) without conflicting with the views you put forward in this thread? I'm curious to know exactly how (and for what) you'd use the output from that position/error analysis? could you elaborate a bit on that?

It's true that Stick was implicitly looking for a better way to measure PR, so I can see why I created the impression that I shared Stick's aims. What I really had in the back of my mind, though, was not how to adjust PR to be more "accurate," but to come up with a model of what kinds of decisions are difficult for human players. This might then be used as the basis for a kind of contempt factor to improve a computer's performance against humans. That is, there might be two moves that are close in equity as we normally think of it, but one play is more complex and increases the chances that the human opponent will blunder, and so should be the preferred play. In my dream world, such a bot would speed up training by steering games towards positions in which I tend to blunder badly, giving me more practice with them.

Messages In This Thread

 

Post Response

Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:

 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.