11-24-2022, 10:24 AM
Hi Franz,
Yes, it's a good sign when the plot goes up, or even flat can be good if the plateau is well above the zero line. The only thing you *really* don't want is a flat line near the zero mark.
The Skill Trace section of the Evaluation dialog (in the next Ludii release) provides the following options:
1. Levels: Number of skill levels tried.
2. Trials: Number of trials per comparison.
3. Time Limit: In seconds.
The process will stop when the time limit is reached or the desired skill level is reached, whichever comes first.
Yes it's interesting that the method also seems to work for games that are more amenable to Alpha Beta-based approaches (e.g. Chess). I think this is because relative playing strength (between successive skill levels) is more important than absolute strength. Also the continuous "any time" nature of UCT means that we get a meaningful result at the cut-off point for each skill level, rather than having to compute complete plies for Alpha Beta-based approaches. Also, the odd-even effect seen with Alpha Beta can throw off the skill level calculation with wildly zig-zagging plots :)
Regards,
Cameron
Yes, it's a good sign when the plot goes up, or even flat can be good if the plateau is well above the zero line. The only thing you *really* don't want is a flat line near the zero mark.
The Skill Trace section of the Evaluation dialog (in the next Ludii release) provides the following options:
1. Levels: Number of skill levels tried.
2. Trials: Number of trials per comparison.
3. Time Limit: In seconds.
The process will stop when the time limit is reached or the desired skill level is reached, whichever comes first.
Yes it's interesting that the method also seems to work for games that are more amenable to Alpha Beta-based approaches (e.g. Chess). I think this is because relative playing strength (between successive skill levels) is more important than absolute strength. Also the continuous "any time" nature of UCT means that we get a meaningful result at the cut-off point for each skill level, rather than having to compute complete plies for Alpha Beta-based approaches. Also, the odd-even effect seen with Alpha Beta can throw off the skill level calculation with wildly zig-zagging plots :)
Regards,
Cameron