View this page "Genetic Algorithm Training"
32 views
Skip to first unread message
Dave
Jun 9, 2010, 8:57:07 PM6/9/10
to Dave Schwartz's World of Horse Racing
Click on http://groups.google.com/group/dave-schwartzs-world-of-horse-racing/web/genetic-algorithm-training
- or copy & paste it into your browser's address bar if that doesn't
work.
Paul
Jun 10, 2010, 12:10:23 AM6/10/10
to Dave Schwartz's World of Horse Racing
Sounds very slick...Darwin would be proud.
One question occurs to me about the training time...what if you had
one, big honking system, running some flavor of Linux, but with a
number of VM's all virtualizing that single core AMD cpu that your
code seems to like best? Is that possible? I don't know if VMWare or
any other product has the ability to create virtual single core AMD
cpu's. Just a thought.
One question occurs to me about the training time...what if you had
one, big honking system, running some flavor of Linux, but with a
number of VM's all virtualizing that single core AMD cpu that your
code seems to like best? Is that possible? I don't know if VMWare or
any other product has the ability to create virtual single core AMD
cpu's. Just a thought.
Dave
Jun 10, 2010, 12:30:58 AM6/10/10
to Dave Schwartz's World of Horse Racing
I am running lickety-split-fine on an AMD 6400 dual-core. The idea is
that it isn't much faster than a 3400-single core.
VMWare will just add a layer in between and slow things down.
that it isn't much faster than a 3400-single core.
VMWare will just add a layer in between and slow things down.
MichaelWilding
Jun 10, 2010, 11:02:16 AM6/10/10
to Dave Schwartz's World of Horse Racing
Very interesting. I was wondering why you chose the top 20% and not
15% or 25% and also the limitations on factors and rules, is this just
because limitations are required for processing or have you found
previously that they are the optimum figures?
Have you ever thought of applying similar AI techniques in order to
determine optimum wagering strategies on contenders based on risk
levels etc...?
15% or 25% and also the limitations on factors and rules, is this just
because limitations are required for processing or have you found
previously that they are the optimum figures?
Have you ever thought of applying similar AI techniques in order to
determine optimum wagering strategies on contenders based on risk
levels etc...?
Dave
Jun 10, 2010, 11:31:53 AM6/10/10
to Dave Schwartz's World of Horse Racing
Michael, please explain your question again. I am not sure where you
are referencing top 20%... If you mean the definition of an "A" horse,
I actually set it at 15%.
Nothing is optimized yet. That has to come later. This is a very
different, brute force type of solution to the problem. IMHO, a
different kind of solution is demanded because the competition has
become so strong - we need to do something that they are not doing.
are referencing top 20%... If you mean the definition of an "A" horse,
I actually set it at 15%.
Nothing is optimized yet. That has to come later. This is a very
different, brute force type of solution to the problem. IMHO, a
different kind of solution is demanded because the competition has
become so strong - we need to do something that they are not doing.
Dave
Jun 11, 2010, 11:23:58 AM6/11/10
to Dave Schwartz's World of Horse Racing
Update: Added some interesting features including ability to train an
ant hill until a certain age (of the best ant) is reached or a certain
accuracy level is reached.
As I watch these guys continue to train with what seems like never-
ending improvement, I am really gaining confidence. I have been
writing GAs and Neural Nets for 20 years and have never seen anything
learn like this.
ant hill until a certain age (of the best ant) is reached or a certain
accuracy level is reached.
As I watch these guys continue to train with what seems like never-
ending improvement, I am really gaining confidence. I have been
writing GAs and Neural Nets for 20 years and have never seen anything
learn like this.
MichaelWilding
Jun 12, 2010, 6:30:54 AM6/12/10
to Dave Schwartz's World of Horse Racing
Sorry for the lack of clearness, I meant in the breeding process. You
say that you only use the top 20% of the ants and I was wondering why
this figure?
When setting the grading levels I assume that this happens once the
ants have been bred to an optimum level. After this would you be using
a brute force method to find the optimum grades? I guess if you are
doing it like this then this means your grade levels will change
depending on the races you are calculating on which would make sense.
say that you only use the top 20% of the ants and I was wondering why
this figure?
When setting the grading levels I assume that this happens once the
ants have been bred to an optimum level. After this would you be using
a brute force method to find the optimum grades? I guess if you are
doing it like this then this means your grade levels will change
depending on the races you are calculating on which would make sense.
Dave
Jun 12, 2010, 10:42:55 AM6/12/10
to Dave Schwartz's World of Horse Racing
I am not sure that you understand the actual process here.
An ant, just like a real human, does not ever change his DNA.
Therefore, his system never changes. The HILL improves through the
breeding process.
Understand that the hill is represented by the best ant. The goal of
the hill is to develop a single ant that is the best example of ant-
hood. That is, the best possible ant he can be.
Thus, at the beginning of a new generation we have these 20 ants left
from the previous generation. They were the best 20 ever produced.
Their offspring replace the 80 ants we killed at the end of the
previous generation. Those were the ones that were not good enough to
make it into the top 20.
Through this process some of the children will be better than their
parents and go on to fame and fortune as breeders themselves, passing
their genes onto the future generations. Ultimately, their parents
will move down the ranked list and out of the breeding pool. When they
do so, they will die. Of course, ultimately, the children and their
children will do the same thing.
As for "Why 20%?" Personal experience. The number of breeders versus
the hill population is not an exact science - at least I know of no
studies that discuss it. It would also depend upon the nature of the
data, the complexity of the system one is trying to uncover. In non-
layman's terms that would be called the "search space."
Consider what we are really trying to do... There are 1,000 factors.
Each factor has a potential of 250 values. There are 500 rules. Each
rule has one of 5 "signs," (i.e. >,>=,=,<,<=), 250 different values, 3
branches (take an action, "and" with the next rule, "or" with the next
rule), and 1,000 possible actions.
That would be 1000*250*500*5*250*3*1000 permutations. Which one is
best? That is the space we are searching to find "optimum." Well, not
really optimum... just VERY, VERY good.
How long do you suppose that it would take to try them all?
Consider the chess board... There are 64 squares x 32 pieces - That is
a HUGE problem. This is magnitudes larger.
An ant, just like a real human, does not ever change his DNA.
Therefore, his system never changes. The HILL improves through the
breeding process.
Understand that the hill is represented by the best ant. The goal of
the hill is to develop a single ant that is the best example of ant-
hood. That is, the best possible ant he can be.
Thus, at the beginning of a new generation we have these 20 ants left
from the previous generation. They were the best 20 ever produced.
Their offspring replace the 80 ants we killed at the end of the
previous generation. Those were the ones that were not good enough to
make it into the top 20.
Through this process some of the children will be better than their
parents and go on to fame and fortune as breeders themselves, passing
their genes onto the future generations. Ultimately, their parents
will move down the ranked list and out of the breeding pool. When they
do so, they will die. Of course, ultimately, the children and their
children will do the same thing.
As for "Why 20%?" Personal experience. The number of breeders versus
the hill population is not an exact science - at least I know of no
studies that discuss it. It would also depend upon the nature of the
data, the complexity of the system one is trying to uncover. In non-
layman's terms that would be called the "search space."
Consider what we are really trying to do... There are 1,000 factors.
Each factor has a potential of 250 values. There are 500 rules. Each
rule has one of 5 "signs," (i.e. >,>=,=,<,<=), 250 different values, 3
branches (take an action, "and" with the next rule, "or" with the next
rule), and 1,000 possible actions.
That would be 1000*250*500*5*250*3*1000 permutations. Which one is
best? That is the space we are searching to find "optimum." Well, not
really optimum... just VERY, VERY good.
How long do you suppose that it would take to try them all?
Consider the chess board... There are 64 squares x 32 pieces - That is
a HUGE problem. This is magnitudes larger.
MichaelWilding
Jun 17, 2010, 4:14:22 AM6/17/10
to Dave Schwartz's World of Horse Racing
Thanks for the answer Dave.
Reply all
Reply to author
Forward
0 new messages