• Hi Guest Just in case you were not aware I wanted to highlight that you can now get a free 7 day trial of Horseracebase here.
    We have a lot of members who are existing users of Horseracebase so help is always available if needed, as well as dedicated section of the fourm here.
    Best Wishes
    AR

Tony's Trainer Model

MySportsAI

MySportsAI

Colt
Tony Spencer recently posted about his trainer model and his work on building trainer behavioural connections. He mentioned that this may or may not be stand alone but could also be an additional input to a wider model
It strike me that a perfect vehicle for this model build would be a Probabilistic Knowledge Graph which in turn would be connected to a LLM
Graphs are an intuitive way of representing trainer behaviour eg think this trainer employed this rider, a graph connection between a trainer and a rider
Connections can not only have direction but also probabilities attached to the connection and furthermore the LLM can utilise these in order to give more nuanced output. Jev also has probability capabilities and may even be a better vehicle although its new kid on the block.
This is an example of a LLM not making predictions in a Machine Learning sense but helping with the heavy lifting of managing and making sense of such a graph. I am putting this here for further thoughts and a nudge should you be interested
 
I might also add that the data we can attach to the graph in terms of connections and nodes is broad for example a trainer to jock connections can contain Trainer SR, Jock SR Same but for this class of race etc etc and of course we are not just restricted to jocks
 
MySportsAI said:
Tony Spencer recently posted about his trainer model and his work on building trainer behavioural connections. He mentioned that this may or may not be stand alone but could also be an additional input to a wider model
It strike me that a perfect vehicle for this model build would be a Probabilistic Knowledge Graph which in turn would be connected to a LLM
Graphs are an intuitive way of representing trainer behaviour eg think this trainer employed this rider, a graph connection between a trainer and a rider
Connections can not only have direction but also probabilities attached to the connection and furthermore the LLM can utilise these in order to give more nuanced output. Jev also has probability capabilities and may even be a better vehicle although its new kid on the block.
This is an example of a LLM not making predictions in a Machine Learning sense but helping with the heavy lifting of managing and making sense of such a graph. I am putting this here for further thoughts and a nudge should you be interested
Click to expand...
That is very interesting and I can certainly see the logic in it.

A lot of what I have been doing with the trainer work is already about trying to move away from looking at a trainer as one single statistic and instead building connections between different behaviours and circumstances.

So rather than saying simply "Trainer X is good", I am looking at things such as trainer/rider combinations, track behaviour, distance, class moves, days since the trainer last ran one, handicap positioning, headgear, apprentices, previous winning marks, seasonal patterns and various other circumstances.

The important part for me has also been chronology. I only want the information that would genuinely have been available before the race, and I am trying to establish whether a particular behaviour is persistent rather than something that has appeared by chance in a small sample.

That is why your Probabilistic Knowledge Graph idea interests me. In effect you could have Trainer → Jockey, Trainer → Course, Trainer → Distance, Trainer → Class Move etc., with the strength and reliability of those relationships attached to the edges rather than treating everything as a simple yes/no connection.

It could potentially go further as well. The interesting information may not be an individual connection but combinations of them. For example, a trainer using a particular jockey at a particular course after a certain preparation pattern might mean considerably more than any one of those factors viewed separately.

At present my machine-learning model is doing much of that combination work numerically. I have a number of individual trainer-behaviour modules feeding a larger model, so I am not sure yet whether a graph would replace that or sit alongside it. My instinct is that it would be more useful as an additional layer rather than throwing away what is already working.

Where I can particularly see the LLM being useful is exactly as you describe — not asking it to "predict the winner", but using it to interrogate and explain the network of evidence. It might be able to say why today's runner looks unusual for that trainer, which historical relationships are supporting it, and which evidence is weak or contradictory.

That would make the output much more interpretable than simply producing a probability.

So yes, definitely food for thought. I hadn't been thinking of the trainer work specifically as a probabilistic knowledge graph, but when you describe it that way there is quite a natural fit with what I am building.

Thanks for the nudge — that's one I'm going to have a proper look at.
 
tony spencer said:
That is very interesting and I can certainly see the logic in it.

A lot of what I have been doing with the trainer work is already about trying to move away from looking at a trainer as one single statistic and instead building connections between different behaviours and circumstances.

So rather than saying simply "Trainer X is good", I am looking at things such as trainer/rider combinations, track behaviour, distance, class moves, days since the trainer last ran one, handicap positioning, headgear, apprentices, previous winning marks, seasonal patterns and various other circumstances.

The important part for me has also been chronology. I only want the information that would genuinely have been available before the race, and I am trying to establish whether a particular behaviour is persistent rather than something that has appeared by chance in a small sample.

That is why your Probabilistic Knowledge Graph idea interests me. In effect you could have Trainer → Jockey, Trainer → Course, Trainer → Distance, Trainer → Class Move etc., with the strength and reliability of those relationships attached to the edges rather than treating everything as a simple yes/no connection.

It could potentially go further as well. The interesting information may not be an individual connection but combinations of them. For example, a trainer using a particular jockey at a particular course after a certain preparation pattern might mean considerably more than any one of those factors viewed separately.

At present my machine-learning model is doing much of that combination work numerically. I have a number of individual trainer-behaviour modules feeding a larger model, so I am not sure yet whether a graph would replace that or sit alongside it. My instinct is that it would be more useful as an additional layer rather than throwing away what is already working.

Where I can particularly see the LLM being useful is exactly as you describe — not asking it to "predict the winner", but using it to interrogate and explain the network of evidence. It might be able to say why today's runner looks unusual for that trainer, which historical relationships are supporting it, and which evidence is weak or contradictory.

That would make the output much more interpretable than simply producing a probability.

So yes, definitely food for thought. I hadn't been thinking of the trainer work specifically as a probabilistic knowledge graph, but when you describe it that way there is quite a natural fit with what I am building.

Thanks for the nudge — that's one I'm going to have a proper look at.
Click to expand...
The LLM does not interrogate the graph your Python code does this, I will see if I can show you a sample tomorrow
 
MySportsAI MySportsAI,
Yes i just gave the data to an LLM to create a graph told you i was thick.
I have now got the LLM to write a python script from data for southwell racecourse Trn, Jky and trainer jockey combined different outcome.
 
You must log in or register to reply here.
Back
Top