Computer-Based Head-to-Head Handicapping

[InControlX]

Originally Posted by Capper Al

Are we saying that the value of each attribute selected increases by the power of 2? Most handicapping factors IVs are in a close range. Yes Pace is better than class, but is it possible that we can be saying 2048 times? I just might be looking at this wrong. Please clarify.

Thanks

Capper Al... No, the binary power number only means how many total combinations of different possible preparation patterns are used. Although some initial binary parameters will have a bigger win/loss or ROI effect than others, we don't need to know this going in. The results run will tell us if our selections were good, i.e., finishing gaps correspond to wins/ROI.

Also, in the determination by final run gaps we find races where a real dog of an entry becomes a pick selection beacuse bad as it's pattern is, the entry is superior to the remaining starters by a good margin. Even so, I still have trouble "pulling the trigger" in these races!

ICX

[DeltaLover]

Originally Posted by podonne

Also, I would not use the standard deviation of the mask's ROI to remove masks. That's only accurate when you use past behavior to monitor future behavior. For partitioning a population into "meaningful" and "not meaningful" sub-groups you need something like a Chi-squared test, or a binomial test since they are binary factors.

As I answered to Al, I think it depends and the correct answer is more of a trial and error process than a clear analytical prove.

Yes ChiSqured is a more accurate test than the one I described and I am extensively using it in other models... It happened that the one I described here is using this method...

[podonne]

Originally Posted by DeltaLover

As I answered to Al, I think it depends and the correct answer is more of a trial and error process than a clear analytical prove.

Yes ChiSqured is a more accurate test than the one I described and I am extensively using it in other models... It happened that the one I described here is using this method...

Fair enough. I'm sure you are using it in the proper context.

Takes me back to the first time I tried something like that. Just seemed so simple. How do I know if a factor is different enough? Just look at a huge number of factors and calculate the ROI's standard deviation, so > 2 dev's is 95% sure its different!

Then someone pointed out that measurements of > 2 deviations were expected in any random process, it was really whether they happened more often than chance would predict (more than 95% of the time). But that just describes whether your sample distribution resembles a normal distribution, not to give you a means of filtering out particular samples, especially without a time dimension of some kind.

Seductive in its simplicity, though...

[DeltaLover]

Originally Posted by podonne

Then someone pointed out that measurements of > 2 deviations were expected in any random process, it was really whether they happened more often than chance would predict (more than 95% of the time).

It took me several iterations to understood this concept.. Of course we can also use a population of totally random factors as a comparison measurement..

[Magister Ludi]

Genetic Algorithms (GA) are frequently used with neural networks. Once a network has been successfully trained to its training parameters, its neurons are "mutated" to create another parent network. The trained network and the mutated network create a child through genetic crossover. The resulting child network is trained and tested. If the resulting network outperforms its parents, its neurons are "mutated" and it is "bred" to create another network. In other words, GA are used only after a network has been successfully trained to create a more accurate and robust solution.

[DeltaLover]

Originally Posted by Magister Ludi

Genetic Algorithms (GA) are frequently used with neural networks. Once a network has been successfully trained to its training parameters, its neurons are "mutated" to create another parent network. The trained network and the mutated network create a child through genetic crossover. The resulting child network is trained and tested. If the resulting network outperforms its parents, its neurons are "mutated" and it is "bred" to create another network. In other words, GA are used only after a network has been successfully trained to create a more accurate and robust solution.

Sure... In the past I have written moduels to train a NN using GA where the weight of each neuron was assigned by the GA avoiding back propagation sigma functions etc... Presently though I no longer implement either one (NN or GA) since I am using Open Source libraries that are making them an implementation detail to the whole platform allowing me to shift my focus to the domain rather than to low level details

[podonne]

Originally Posted by Magister Ludi

Genetic Algorithms (GA) are frequently used with neural networks. Once a network has been successfully trained to its training parameters, its neurons are "mutated" to create another parent network. The trained network and the mutated network create a child through genetic crossover. The resulting child network is trained and tested. If the resulting network outperforms its parents, its neurons are "mutated" and it is "bred" to create another network. In other words, GA are used only after a network has been successfully trained to create a more accurate and robust solution.

True, but there are a ton more applications than just fine tuning neural networks. Its really powerful when faced with a huge range of solutions where you can develop a bunch of random solutions, pick the best, and combine them to create a bunch more solutions, pick the best, and repeat.

Not sure you meant to be restrictive, just didn't want to leave the impression that application was its purpose.

For fun reading along the lines of this thread take a read about Learning Classifier Systems (LCS). Genetic algorithms + reinforcement learning + masks. Fun stuff. http://en.wikipedia.org/wiki/Learning_classifier_system