Pattern Recognition

[traynor]

Originally Posted by Robert Fischer

The "Blink" idea of PR is very much related to the idea of intuition. similar = RPD http://en.wikipedia.org/wiki/Recogni...rimed_decision

This one of the most popular (and most useful) schemas for managerial decison making in business. It is an ongoing, continual process of calibration to continually improve the quality of the decisions. Military loves it, too.

http://en.wikipedia.org/wiki/Sensemaking

[DeltaLover]

Originally Posted by traynor

That is the entire point. It is not that difficult. It is as doable, attainable, accessible for the average bettor as for the average graduate student. The major advantage the average graduate student has is that he or she understands that she or he does not have all the answers already, and that some serious study is going to be needed to do well.

It is not necessary to learn multivariate calculus to train pattern recognition skills. It helps if the person designing the training understands it, though. As well as machine learning, data mining, artificial intelligence, and a number of other topics. Large doses of cognitive psychology and learning theory help, too. Otherwise the "training" is considerably less useful.

What is necessary (or at least very useful) is to realize that there is a great deal more to "pattern recognition" than the pop culture view.

[Dave Schwartz]

Originally Posted by DeltaLover

Delta, I love this. You actually resurrected a 5-year old thread with a .

That's priceless.


Then I saw that one of the last posters was Bobby Goren. Added a touch of sad.


Dave

[DeltaLover]

Originally Posted by Dave Schwartz

Delta, I love this. You actually resurrected a 5-year old thread with a .

That's priceless.


Then I saw that one of the last posters was Bobby Goren. Added a touch of sad.


Dave

Yes Dave!

5-year old but still extremely important!

Sad thing though is that today people do not care about this kind of discussions!

[Gamblor]

So... you guys managed to mechanise pattern recognition yet?

[classhandicapper]

Originally Posted by traynor

In particular, Decision Traps, which has a very nice section devoted to a study of professional thoroughbred horse race analysts, in which the researchers discovered that the fewer pieces of information the analysts used, the more accurate their predictions were. In contrast, the more pieces of information they used, the less accurate their predictions became, but the greater their confidence in the accuracy of those predictions. So much for the use of untrained pattern recognition skills. It also does not speak well for "comprehensive handicapping."

I think there are multiple problems with too much information.

1. A lot of the information we use is not actually accurate. Whether we are evaluating horses using pace and final time figures or comparative class and race flow we are always working with some flawed information

2. When you combine multiple sources of flawed data, it sometimes introduces new flaws or compounds the ones already there.

3. When you use a lot of information it gets harder and harder to weight the different pieces of information against each other.

4. Combining pieces of information can sometimes increases accuracy but reduce value.

These are the kind of issues I've run into for years when trying to combine my personal classing insights with pace and speed figures.

The problem of course is that in horse racing ignorance is not bliss either.

[DeltaLover]

Originally Posted by classhandicapper

I think there are multiple problems with too much information.

1. A lot of the information we use is not actually accurate. Whether we are evaluating horses using pace and final time figures or comparative class and race flow we are always working with some flawed information

2. When you combine multiple sources of flawed data, it sometimes introduces new flaws or compounds the ones already there.

3. When you use a lot of information it gets harder and harder to weight the different pieces of information against each other.

4. Combining pieces of information can sometimes increases accuracy but reduce value.

These are the kind of issues I've run into for years when trying to combine my personal classing insights with pace and speed figures.

The problem of course is that in horse racing ignorance is not bliss either.

What you are mentioning here, consist well known issues for any kind of data mining regardless of the particular domain that is modeled.

We must start with as simple and primitive data and gradually add new features until we reach the desired outcome; feature engineering is the most challenging task in the machine learning pipeline as it not only requires advanced understanding of the underlying theory but a great degree of intuition and related experience from other niches as well.

Procedures like reducing the dimensionality of the input pattern, picking the most applicable normalization methodologies and applying several learning methodologies until an acceptable solution is found can not be addressed by a non expert in the field of Data mining and machine learning. More than this, I am sure that the layman can not even articulate or understand the real problem or the main questions that need to be answered to successfully develop such a system. Of course this difficulty is a great thing as it keeps the game interesting and challenging; if it was simple to solve, horse betting simply would have been another form of an unbeatable casino like game.

[classhandicapper]

I'm most interested in building automated procedures that spit out races and horses that fit the parameters of what I might normally look for in a potential bet when I handicap races thoroughly. That way, instead of handicapping 9 races thoroughly and sometimes finding little or nothing, my procedures can quickly look at many race cards and identify 5-6 races and/or 4-5 horses that may be fertile ground for a bet.

What I've found is that I can't program as well as I can think because the possibilities are almost endless and I want to be able to use some subjective analysis and common sense. But what I can do is program parameters as a shortcut. Then I can look at the specifics and see whether that race/horse is actually as fertile grounds for a bet as I hope.

The other thing I do is research.

Formerly, if I had some new theory or I read something new in a book it would takes years of experience or weeks of manual labor to research it. Now I can write a query, run it against years of data, and separate the BS that loads of people believe because it was written by some reputable author or they had some short term positive success from the stuff that is actually correct.

[DeltaLover]

Originally Posted by classhandicapper

I'm most interested in building automated procedures that spit out races and horses that fit the parameters of what I might normally look for in a potential bet when I handicap races thoroughly. That way, instead of handicapping 9 races thoroughly and sometimes finding little or nothing, my procedures can quickly look at many race cards and identify 5-6 races and/or 4-5 horses that may be fertile ground for a bet.

This problem is relatively easy to solve. I can see two obvious solutions:

(1) Use a supervised learning approach. You will need to create a labeled sample consisting of the representation of the races and the corresponding labels (playable - not playable) (1k races should be enough). You can start training a logistic regression model and if this does not solve your problem you can try a simple perceptron based neural network.

(2) Another solution might be to use k-means clustering to build the groups that fit your rules. This approach will be more difficult as you will need to work hard in the data normalization and possibly engineer composite features to allow for valid euclidean distance comparisons.

[Robert Fischer]

Originally Posted by Gamblor

So... you guys managed to mechanise pattern recognition yet?

I have no idea. I do know I could assist in programming a good one.

There are also at least a couple significant different uses/perspectives of the idea of 'Pattern Recognition';

A) prototypical situations, which call for a specific weighting of models within a heuristic search, aiming to find a horse being offered at a value

B) prototypical situations where the public is significantly likely to make mistakes/biases

[classhandicapper]

Originally Posted by DeltaLover

This problem is relatively easy to solve. I can see two obvious solutions:

(1) Use a supervised learning approach. You will need to create a labeled sample consisting of the representation of the races and the corresponding labels (playable - not playable) (1k races should be enough).

I already have several very good working models up and running spitting out reports, but my list of ideas and ways to make them better grows faster than my ability crank out code.

[ReplayRandall]

Originally Posted by DeltaLover

This approach will be more difficult as you will need to work hard in the data normalization and possibly engineer composite features to allow for valid euclidean distance comparisons.

Delta, couldn't you have just said "weak ordering" instead of having to inject another reference to your favorite Greek mathematician??...

[DeltaLover]

Originally Posted by ReplayRandall

Delta, couldn't you have just said "weak ordering" instead of having to inject another reference to your favorite Greek mathematician??...

See here:

https://en.wikipedia.org/wiki/Euclidean_distance
[url]https://en.wikipedia.org/wiki/K-means_clustering[/url

(There exist alternatives like Manhattan or Hamming distances but Euclidean is the "standard" solution).

[ReplayRandall]

Originally Posted by DeltaLover

See here:

https://en.wikipedia.org/wiki/Euclidean_distance
[url]https://en.wikipedia.org/wiki/K-means_clustering[/url

(There exist alternatives like Manhattan or Hamming distances but Euclidean is the "standard" solution).

From your same Wiki source:

"The points of the Euclidean plane may be ordered by their distance from the origin, giving another example of a weak ordering with infinitely many elements, infinitely many subsets of tied elements (the sets of points that belong to a common circle centered at the origin), and infinitely many points within these subsets. Although this ordering has a smallest element (the origin itself), it does not have any second-smallest elements, nor any largest element."

[DeltaLover]

Originally Posted by ReplayRandall

From your same Wiki source:

"The points of the Euclidean plane may be ordered by their distance from the origin, giving another example of a weak ordering with infinitely many elements, infinitely many subsets of tied elements (the sets of points that belong to a common circle centered at the origin), and infinitely many points within these subsets. Although this ordering has a smallest element (the origin itself), it does not have any second-smallest elements, nor any largest element."

definetly!