In the past both features and their parameters where handcrafted.Sergei S. Markoff wrote: I'm disagree. Most strong programs are tuned using huge position/games datasets, eval function is complicated and has a lot of params, at least several thousands.
In other hand a neural network is nothing more then hierarchy of logistic (or other activation) functions. I don't see any significant difference, wouldn't you?..
With the introduction of powerful testing frameworks, the feature parameters started to be machine tuned: this has been a big improvement.
Now Giraffe has proved that is possible to go to the next level: even the features themselve can be automatically extracted out of row data.
I think Giraffe is the first experiment to really prove this is possible, other experiments that failed to produce a good working engine are to be ignored (only successful results make experiment interesting for me; this heuristic allows to filter a lot of garbage).
Differently from parameter auto-tuning, the NN approach has still to prove that is stronger than traditional one. IMO this is a difficult task to win because chess is intrinsically a game more suited for computers than for humans (I don't know about "go").