Hacker News new | past | comments | ask | show | jobs | submit login

Deep Blue isn't ML at all, it's just a purpose built system for solving chess. It doesn't learn anything as such and just a fancy brute forcing machine that (smartly) goes through all possible chess moves and selects the best possible tree of outcomes. The work on actual ML happened independently of this.



That is a form of machine learning. It uses an algorithm with an enormous ability to look ahead, and selects the moves according to a heuristic that have been shown previously to be most likely to create the desired end state.

Chess is not a solved game from the starting state, so it has to make assumptions based on the data it has. This is machine learning. Please don't make such definite statements. You even say it in your own comment "selects the best possible true of outcomes", how do you think it selects this? It uses heuristics to assign values to different board states, and in the case of Deep Blue these values were created through previous game analysis. If a knight to c2 on turn 8 is rarely seen in the same game as a winning board state, then this is valued lower. Looking through the tree wasn't the ML part, but knowing how to pick the best node on the tree was. Deep Blue is ML.


> It uses heuristics to assign values to different board states, and in the case of Deep Blue these values were created through previous game analysis.

Unless I’m misinformed, this part isn’t true. The heuristic was hardcoded with the help of human experts.


From the paper Deep Blue by Murray Campbell et al (people who worked on Deep Blue) "The initialization of the feature values is done by the “evaluation function generator”, a sub-program which was run on the master node of SP system." Which would suggest it generated the heuristic itself. The features may have been hardcoded, but assigning values to them wasn't. In addition to this, feature values could be static or dynamic, meaning it would update dynamic ones depending on the board state. It not only generated the heuristic values (feature values) it could modify them to reflect their relative change in impact throughout the game.


mehh feels like we are creating fancy ML brute force crunching machines now


Now = last 40 years




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: