On Fri, Jul 18, 2008 at 10:52 PM, Michael Williams
<[EMAIL PROTECTED]> wrote:
> I like the concept of Green Go, and the use of EC2, but your footnote is
> very confusing. Especially the last sentence.
okay, that's this part:
> Ryan Grant wrote:
>> [1] for each game pairing, the game-relative joules is the total
>> difference in energy used between opponents. it's positive for
>> one player, and negative for the other, unless both use the
>> same energy - then it's 0 for both. it's independent of win or
>> loss. if your computing architecture requires higher wattage
>> than the opponent, resign lost games quickly. joules used for
>> pass-moves should be free.
although Green Computer Go creates a "new way to win", it should
always aim to reward better Go play, with a bonus to algorithms
that do so using less transistor switching. incentives for
strange playing styles should be minimized.
one can imagine some incentive mistakes that could arise. for
instance, if the tournament measured each contestant's total
energy used across all games, a player could fall behind in the
tournament by initiating a long ko, even if they ended up winning
the game.
the quoted footnote tries to explain that in a single game, only
the difference in energy usage is counted. if, at the end of the
game, player A used 1300 joules, and player B used 800J, then
player A must add 500J to their total, while player B gets to
subtract 500J. if, in another game, players A and B both use
900J, then the difference is 0J, and their totals are unchanged.
if one is counting joules, then one likely knows how many joules
were used in each move. the footnote suggests that energy
required to come up with a pass should not be counted in the
total energy used by a contestant for that game. in normal
games, this will have no effect. but it removes an advantage
that a lower-power opponent might use to prolong a lost game,
thus creating more of an energy difference. more on that below.
here are current goals of the energy monitoring plan:
* do not penalize some contestants for playing more games in
the tournament than others, since this is outside their
control.
(in the above scheme, a power-hungry contestant may do
progressively worse the more games they play, but at least a
contestant using a normal amount of energy has a fair chance,
no matter how many games they play.)
* do not penalize contestants for continuing long games.
(it may be necessary, to play better Go.)
* minimize creating a new incentive for contestants with lost
positions to play moves that cannot change the outcome of the
game.
(it's always legal to keep playing, and moves create time
consequences for the opponent, but time differences do not
carry forward into the next game, while the energy difference
currently does. therefore, the energy consequence for
responding to questionable moves should be mitigated in some
way. my current proposal is to not count the energy used to
choose a pass-move. this assumes that there is
never-or-rarely an advantage to a pass-move.)
with energy consumption introduced as a second variable into
tournament results - and with it not correlating 100% with
winning on the board - there will have to be some time to settle
out the incentive bugs.
i suspect one area requiring further thought will be minimizing
the effect of losing to an army of weaker bots that require
slightly less power. this suggests that energy consumption
should be averaged and weighted according to how close the
ranking of the opponent is to one's final rank.
a last feature which would be fun to introduce, for the humans,
is some escalation of importance of the rounds, reaching a
climax.
- Ryan
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
