On 21/03/2017 21:08, David Ongaro wrote: >> But how would you fix it? Isn't that you'd need to retrain your value >> network from the scratch? > > I would think so as well. But I some months ago I already made a > proposal in this list to mitigate that problem: instead of training a > different value network for each Komi, add a “Komi adjustment” value as > input during the training phase. That should be much more effective, > since the “win/lost” evaluation shouldn’t change for many (most?) > positions for small adjustments but the resulting value network (when > trained for different Komi adjustments) has a much greater range of > applicability.
The problem is not the training of the network itself (~2-4 weeks of letting a program someone else wrote run in the background, easiest thing ever in computer go), or whether you use a komi input or a separate network, the problem is getting data for the different komi values. Note that if getting data is not a problem, then a separate network would perform better than your proposal. -- GCP _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
