the main reason why i am inerested in this line of reasoning is that even though we hvae learned to make programs that play brilliant chess the programs themselves are unable to communicate their thought process to us. it is hidden in the various convolutions and angles of the particular search space that they search through and this is incomprehensible to any human being. what we need is intelligent search and programs that use data mining to show us the way forward.
using machine learning/data mining/cognitive science to make intelligent and responsive chess databa
Sort:
i just came up on the following. those who have read my thread on machine learning have some background if you have not read that thread i suggest you first do so.
let's say there is a person let's call him Andrew. He enjoys water sports but only in certain conditions. let's say the parameters are temperature, humidity and wind speed. now if it is cold and humid and win speed is normal then he enjoys the water sports. this is a training example. how do we design an algorithm to predict when andrew enjoys the water sport. a hypothesis function must be created to map the weather conditions to whether or not he enjoys the water sport. we start with the most specific case and then use the training examples to generalize. first we start with the hypothesis that andrew does not enjoy water sports whether the temperature is hot or cold, the humidity is high or low or the wind speed low,normal or fast. we then test against the first example. suppose the first example tells us that he enjoys the water sport for temperature=low, humidity=high and wind speed=normal. then our new generalized hypothesis will be that he enjoys the water sport for temperature=low, humidity=high and wind speed=normal. we then test against a new training example. suppose this time all else is same but humidity is low. now our new hypothesis will be that he enjoys the water sport for temperature=low, humidity=? and wind speed=normal. where '?' represents 'any value'. as we train more and more we keep coming up with more and more general hypothesis. as we apply this hypothesis to a particular set of variables we come up with a prediction for whether andrew enjoys the water sport on that particular occasion. we then have 'learned' to make this prediction.
now i am thinking that we can use this sort of algorithm and apply it to databases to figure out when certain kinds of sacrifices are successful. so for example, when our bishop is on d3, our queen one move away from the h-file and a knight one jump away from g5. now we consider the case when the opponent has a knight on f6 and whether it is pinned or it is not pinned and the second case that he does not have a knight on f6. another case would be when he does not hvae a knight on f6 but he has a knight which is one jump away from f6 so that it can come to the defense of the king one move after the sac is made and mate is threatened. we can apply this kind of specific to general hypothesis learning model to figure out when these sacrifices are successful. naturally he may have a knight on f8 etc. or other ways to defend. but i'm thinking we can mine the database and get useful information on typical sacrifices.