DJ Teflon wrote:Let me give an overview:
The Algorythm is awesome.
If it reflected the value of a win per se as well as everything it already does it would be even more awesome.
A valid concern. Actually the algorithm does take this into account.
This is what it looks like on my page:
Min Delta per Win/Loss [value = 200] Win Rate of 50.001% adds/subtracts this value from opponent's score
Max Delta per Win/Loss [value = 600] Win Rate of 100% adds/subtracts this value from opponent's score. For values inbetween, Delta = Min + 2*(Win Rate%-50%)*(Max-Min)
By default, I have min at 200 and max at 600. When you load the page, it will always have these values, but you can edit them on the page. This means that the least amount of additional points you get from a win is 200 (even if you barely win), and it can go up to 600 depending on the win margin.
If you have min set to 0, then it will be like how you thought it was (i.e. barely winning is nearly equivalent to barely losing). I tried it out with both values and decided the results looked better with min = 200.
There are two ways you can decide what to use for values here, those that maximize model accuracy and those that seem right. The ideal way would be to measure model accuracy given every combination of min/max values (just keep max at 600 and vary min would work). Then choose the value of min which resulted in the highest model accuracy. I think there wasn't enough data at the time for a comparison of model accuracy to be compelling either way.
Model Accuracy calculation is built in, btw.