>> Typically, how many parameters do you have to tune ? Real or two-level ?
>
> I guess I have 10 real valued and 10 binary ones. There are probably a lot 
> of stuff that are ahrd coded and could be parameterized.
>
> Here I am also completely ignoring playouts that have hundreds of handtuned 
> parameters.

There automatic techniques like those describes by Rémi Coulom are
probably useful.

>> If you consider a yet reasonable subset of parameters, an efficient way
>> to estimate them is  to use fractional factorial design for the linear
>> part, and central composite design for quadratic part (once you know you
>> are already in the right area). You are much more precise than with
>> change-one-at-a-time strategies if there is no interaction between
>> parameters, and you can detect interactions.
>
> I once met this guy:
>
> http://meche.mit.edu/people/faculty/index.html?id=27
>
> his research is a mix of testing formal methods and how well they work in 
> practice and also studying how engineers (who often do not use these 
> methods) actually do in practice.
>
> He seemed to argue that doing parameter optimization intuitively by hand is 
> not as bad as one might think compared to fractional factorial design. So I 
> use that as an excuse for just doing it as I always did. For me it is 
> important to keep a careful record of what I do and plot the result with 
> confidence intervals to avoid tricking myself.

Anyhow, it's always wiser for someone to use a method he understands; so
best to keep it simple if you do not take/have the time to learn more
sophisticated techniques.

>> Alternatively, especially with a very high number of real parameters,
>> derivatives of MC techniques can be efficient and easy to implement:
>> particle filtering or swarm optimization in particular.
>
> That would be tempting (I once implemented a fitting method inspired by 
> simulating annealing and it was very efficient) but it would require a 
> completely different test setup than the one I use right now.
>
Thinking more about it, that's not completely obvious, and hence is
interesting (for me): usually those methods are tailored for functions where 
they get the exact value, not a Bernoulli trial. 

> It also a matter of time and patience. I want new results every day. If I 
> would test all parameters at once using formal methods I would still have 
> to wait for weeks.

Well, once you are in the right zone, if you want to check a change, you
can look at this change only.
Note that you can also try and get information on other real parameters
even when focusing on a precise one: you make them vary very little.
Then the results for those supplementary parameters nearly do not impact
your main parameter, and you can get information through taking means
for a longer time...

Jonas
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to