Can't sign in? Did you forget your password?
If you don’t have an account, click right here
Can't Sign in? Did you forget your password?
to create your profile
Click on the image to upload your profile pic
Enter your email and we shall mail you the password reset link
Enter new password and confirm password
Here we discuss, well, everything science! That includes research, new discoveries, applications of science in the world, science policy, Technology and innovation
But to get to the bottom of your question, you can think about it like this: a recursion model will attempt to know every possible state and solution to a problem. Show the model any given chess board and it will be able to look into its memory, consider all possible next moves, and choose one that yields the most probable path to victory.
It seems pretty straight forward, that's what you would want a chess playing algorithm to do. The problem is that chess is a game with a lot of entropy (complexity) and so it's very inefficient to have the machine remember all possible game states. What would be better is if the algorithm could generalize, taking one chess board configuration from its memory and learning a lesson from it that it can apply to different scenarios.
Now imagine instead that each page of the book had strategies, patterns you can recognize that apply to many different situation. The book could be so much less than 5800 pages and would still be just as effective! That is basically modern machine learning, the model can generalize strategies.
Wth a game like chess and go there are way more than 5800 possible moves so it is very important we get that page count down, our computers can only remember so much!
Backtracking algorithms work by searching through millions of different possibilities for what could happen next in the game, trying to find some way to force their opponent into losing. Deep neural nets don't do this. Instead they play like José Raúl Capablanca and don't do any searching at all. Instead, they look at the entire board and use sophisticated pattern matching to guess if it looks more like the kind of game that's likely to lead to a win vs. a loss. Often this pattern matching is based on a huge database of professional games, combined with a long pre-computation to fine-tune the patterns by having the computer play itself millions of times and seeing how often the neural net's guess is correct (and slightly modifying the neural net accordingly).