On Fri, Aug 21, 2009 at 17:29, Frank Berger
<address@hidden> wrote:
When youy train the NN< it is trained. Whether your training INTENDS to
train it in that way does not change whether or not all or a portion of that
training is in fact DOEN by the training you select for another purpose.
So your arguemnt is again specious.
It may be specious for you but for me it's given that predicting MT is difficult from the dice alone (naturally if you know it is a MT and if you have a sequence you could find where you are, but a NN can't do that).
Again, specious.
The NN + DB _can_ do that TOWARDS THE GOAL AS IMPLEMENTED: to win.
it doe snot need all that side trapping you add in.
it just seeks TO WIN.
If it _can_ learn (when the DB is "open"
and to a more limited extent against static state space folding of a dynamic sequence
thatis inherent in earlier learning):
It _can_ do that.
That an NN that should learn something BG should do learn to exploit irregularities in the RNG is to me a very remote possibility.
The NN+DB does not care waht is remote to you.
At ealry learning time ALL sequences are equally remote.
I don't know whether you know how a NN for BG is trained, but as I pointed out that in the case of TD(Lambda) the net sees only the resulting positions, not the dice. If the effect you suggest is so
it sees the ful stae space IMPLIED and inhrentin those positions.
but YOU do not see THAT.
obvious just demonstrate it. I will happily send you a box of wine if you succeed. All what you suggest makes only sense if the RNG are crap and I don't believe in that.
the RNG are fine, but the definition of "rnadom" is a very slippery one.
You have to go back to what the original definition of a function is:
when you realize you do not even have to be able to NAME the numbers involved or have ANY element of predictabiltiyi n clasical sense for it to be predictable in a NON classical sense, you will start to get the idea.
Draw me a full Pinero curve,pleae.
Can't be done,.
yet the function is a real one.
And has properties that are very detectable and utilizable.
What I do know is that the influence of the random initialization of the net at the start when using TD(Lambda) could easily be in the area of 0.005 in cubeless play. I would bet that a net trained to predict MT would fail to have 0.005 better prediction than random prediction (a long enough sequence for sure). Again I will happily send you a box of wine if you succeed in prediction of MT
You miss the pit: if there is ANY variation off ideal probability
YOU have to explain it.
In the gnubg training process, AFAIK the net is trained to approximate rollout results from a large number of positions. If I'm right with that, there would be no predictable influence of the RNG at all.
Not posiitons. Sequences of play arriving at positions. Evne wereit only positions,
the state space implie all information inherent at ARRRUIVING at those postitions
relative to the rules of BG, and such
carries a portion of the eignestate of the sequences that arrive at those positions.
And finally, no bot I'm aware of adapts when delivered (for obvious reasons), so there is no cheating. There *is* the possibility to analyze the game of the opponent and to exploit his errors/style of play, but this wouldn't be cheating and would only weaken the game play overall.
As I defined the term "cheating", yes, there is.
but you will not consistently use the definitions as given:
yuo arbitrarily track off to your own withou regard to communicational consistency.
Shift gears into math and try to say what you have said in math.
You will start to see your inconcistency.
it still does not mean you wil agree with what I said
or that all of it is correct.
But you wil start to flush out some gross (large) consistency errors in your reasoning.
ciao
Frank