On 6 January 2012 10:03, Philippe Michel <address@hidden>
Ok, so a worthwhile improvement would be a decrease of at least 10% in benchmark error, something like that ?
On Thu, 5 Jan 2012, Joseph Heled wrote:
Seems like you got a very very slightly better race net, but I would be
surprised if it makes a difference in real life.
That would be nice :)
What do you mean by roll-out database ? The benchmark data, since the training data seem to be 2ply evaluations ? (at least the comment in crashed-train-data says so). Or is it both ?
Would be much more interesting to
- get a better contact or crashed net
- expand the roll-out database for all categories (should be easy with
the current availability of cycles)
"roll-out database" == "benchmark data"
If training data is indeed 2ply evaluations, it would be relatively easy to use 4ply instead. And maybe rollouts for those where 2ply and 4ply (or 3ply and 4ply) differ most, but would, say, a few hundred rollouts with accurate equities among hundred of thousands of evaluations be useful or possibly counterproductive singularities ?
I doubt 4 ply would make a difference, but what do I know. But a combination of 2 and 3 might be interesting. I am sure better training schemes are possible, mine just evolved. One point is that the benchmark is not used in training, so is can be used as an "impartial" arbiter between nets. But the benchmark is only a sample, and so, the more comprehensive it is, the better. The idea of adding duplicates is intriguing.
>From the comments in the files, it seems you added positions in batches of a few thousands. You couldn't add so many specific positions, could you ? How were they chosen ? You start from an "interesting" one and add the positions after the few best moves for the 21 possible rolls for, giving, say, one hundred related positions, something like that ?
Not sure what you mean by "specific positions". Positions are usually added in pairs, to try and get the bot to make better move decisions. you should study the code carefully for the exact details, but my choices are certainly sub-optimal.
There are no duplicates in the training data. Isn't it a concern that the recurring early game positions are drowned among much less common ones ?