| |
BGonline.org Forums
Testing empirically backgammon rules of thumbs
Posted By: 6stones In Response To: Experts disagreed on this one (Stick)
Date: Friday, 27 July 2012, at 2:03 p.m.
Fair enough, I stand corrected (I guess), and good to learn, thank you. Would you, Mr. Rice and Mr. Koca, happen to have a particular score in mind which you could mention off the top your head?
I was applying an unvalidated hypothesis, a hasty generalization ("safer alternative play at GG tend to be correct at other scores as well"), and what is more, I am painfully aware how imprecise and idisyncratic the definition of bold/safe is, e.g. one could well argue that making the 5-point here is to be considered the safer alternative.
Applying hypotheses like this is similar to several rules of thumbs, or heuristics, commonly used when playing the game ("Make the 5-point!"). As is well-known, heuristics don't guarantee the correct solution in all cases but work decently enough to be useful, they ignore some context-specific information to save our limited computational resources and time, for better or worse, they allow one to make a quick decision without analysing the position any deeper. (They are like context-free rules discussed in another (sub-)thread on Dreyfuss model of expertise)
I haven't thought so much about the details, but one of my hopes and wishes actually would be to take the time in the near feature and try to empirically test the validity of several common rules of the thumbs in backgammon by gathering hundreds or thousands of positions which are relevant for a given heuristic, and study how they fare in XG rollouts when varying different parameters (specificity and scope of heuristic, decision types, game phase and position type, match score, cube value and location, etc.) Obviously, the more specific the rule of thumb, the more it pays attention to context-specific features of the position, the more processing resources and time it would take. What would be sets consisting of dozens of backgammon heuristics like that would be ideal tradeoffs for different purposes in terms of accuracy and cognitive load/speed? How would different versions of a bot program that can only resort only to limited number of heuristics fare when varying the set of rules of thumb? Are there useful (meta)heuristics for resolving cases of conflicting rules of thumb? Under which circumstances does resorting to a heuristic become counter-productive? (In fact, an expert human player might be tempted to dismiss any general principles and judge any position and decision on its merits, although I am not at all sure to what extent they in practice are able to dispense with them altogether.)
It remains to be seen how much of this positional data gathering can be automatised. There is a lot of routine work but hopefully at least I would learn more about backgammon and the nature of heuristics, both of which I am currently interested in, and perhaps in the long run slightly increase our understanding of heuristics: as of now the scientific study of heuristics both as the general art/science of problem solving (in George Polya's sense) and as methods/mental shortcuts of (semi-)intuitive decision making in various contexts is pretty much on its infancy - the work done in the fields of artificial intelligence and computer science with their definition of heuristics as a class of those methods which are opposed to algorithmic is of course somewhat more developed and at least partly of relevance to the abovementioned.
Perhaps someone else would care to comment on what would count as a scientifically rigorous testing in this case.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.