Hi Hiroshi,

I think the problem with adding moves of the opponent is that they
don't come from the same distribution. E.g., imagine if Black is
winning 90% then it's normal rave moves are also winning 90% on
average (so mixing with the direct statistics doesn't cause a big
change on average regardless of the mixing ratio). However, once you
start to add White's moves the distribution is reversed (on average
only winning 10%) and together those statistics will average roughly
on 50%, which causes a systematic bias.

The intersections where both sides want to play first will of course
still stand out, so that's why reversing the sign helps at least a
bit. However, to get this to work really well you need to find a
clever way to get the bias down... Expand on this a bit and you have
yourself a new rave paper ;-)

Best,
Erik


On Fri, Aug 12, 2011 at 7:18 PM, Hiroshi Yamashita <[email protected]> wrote:
> Hi Erik,
>
>> I think the usual procedure is to update only moves by the same
>> player. However, if you also want to use the opponent's moves (which
>> is not an unreasonable idea), then you should probably reverse their
>> sign (because those moves lost), so you only get something like:
>
> I tried this idea.
>
> 9x9 (1000games 6000po against Fuego 10000po)
> 0.516 RAVE normal (same player only)
> 0.211 RAVE Black and White
> 0.264 RAVE Black and White,      reverse their sign
> 0.092 RAVE not same player only, reverse their sign
>
> 19x19 (1000games 700po agaist GnuGo Lv0)
> 0.570 RAVE normal (same player only)
> 0.013 RAVE Black and White
> 0.078 RAVE Black and White,      reverse their sign
> 0.141 RAVE not same player only, reverse their sign
>
> Normal is clearly better, but "reverse their sign" looks a bit meaningful.
>
>> In a B node, update only B moves - and only if there's no W move at
>>  same location before B's move.
>
> I use this in 9x9, but I ignore all W moves in 19x19. I got a little
> better result from this.
>
> Regards,
> Hiroshi Yamashita
>
>
>
> ----- Original Message ----- From: "Erik van der Werf"
> <[email protected]>
> To: <[email protected]>
> Sent: Thursday, August 11, 2011 8:42 PM
> Subject: Re: [Computer-go] RAVE uses all moves in playout?
>
>
> On Thu, Aug 11, 2011 at 3:43 AM, Hiroshi Yamashita <[email protected]> wrote:
>>
>> Hi,
>>
>> I have a question about RAVE.
>>
>> For example, there are 5 moves a,b,c,d,e. Black to play.
>> In playout, Black plays "a", White plays "b", Black plays "c",
>> W plays "d" and B plays "e".
>> Then game is over, result is B win.
>>
>> B:a W:b B:c W:d B:e ... B win, result = +1
>>
>> I update only B moves RAVE, like this.
>>
>> RAVEcount(a) += 1, RAVEwins(a) += 1
>> RAVEcount(c) += 1, RAVEwins(c) += 1
>> RAVEcount(e) += 1, RAVEwins(e) += 1
>>
>> But on Sylvain's paper Figure4,
>> It looks like updating all moves Black and White, like this.
>>
>> RAVEcount(a) += 1, RAVEwins(a) += 1
>> RAVEcount(c) += 1, RAVEwins(c) += 1
>> RAVEcount(e) += 1, RAVEwins(e) += 1
>> RAVEcount(b) += 1, RAVEwins(b) += 1
>> RAVEcount(d) += 1, RAVEwins(d) += 1
>>
>> Is this right RAVE?
>> I have misunderstood long time.
>
> Hi Hiroshi,
>
> I think the usual procedure is to update only moves by the same
> player. However, if you also want to use the opponent's moves (which
> is not an unreasonable idea), then you should probably reverse their
> sign (because those moves lost), so you only get something like:
>
>> RAVEcount(b) += 1
>> RAVEcount(d) += 1
>
> Best,
> Erik
> _______________________________________________
> Computer-go mailing list
> [email protected]
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
> _______________________________________________
> Computer-go mailing list
> [email protected]
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to