This comment is great, LMFAO!
Machine Learning and statistics
- Thread starter justnum
- Start date
Aka: you remain full of something that smells bad and I wouldn't want to step in.
What does this post have to do with compiling a database of pool match results on which to build a model?
Markov process setting up weights. finding eq (equilibrium) solution.
some people think everything is binomial distribution.
throw in stochastic just in case anyone was in economics.
for me I just program each one.
random events -> model
Given there are an infinite number of options after the break I think the important part would be to determine distance between balls, clusters, how many times does one need to cross the table to get from one ball to another, how many balls are limiting the routes the CB could take, etc.
Clearly, the images would be needed to determine an algorithm but I don't think the images would be used to determine the outcome.
That could be fun training the machine to learn how to play safes in a variety of situations. that aligns with another project im working on.
I was referring more to the ability to run a table after the break - statistically based on their skill level.
I.e. I break and the computer tells the audience I have a 50% chance of running the table if I go this way, or 65% chance if I go that way and it changes with each shot based on the next leave. It could also say "HAL recommends a safety shot after pocketing the 3B", etc.
Clearly, we can watch good players and get a feeling if they'll run out or if we think they SHOULD run out, but what if it could record us and determine that when a ball is 3" off the rail at X angle and X distance from the pocket I make that ball 98% of the time, then when whitey settles down for the next shot it determines the ball is 12" off one rail, 18" off the other, with X angle for the CB and I make that ball 92% of the time.
That would cool.
Can't really be determined. Execution accuracy is too much of a factor.
Thrilling. I am sure we all expect you to follow through on it.
We've been working on something since the start of this year that would use machine learning and object tracking to analyze a video of a pool match and generate the correct data that a third party manual inputs into our Pool Stats Pro app. The bottleneck is the break, identifying stripes versus solids for where the camera may be located, among other things like how it will know what is a defensive shot versus a missed shot. There are other bottlenecks to have machine learning input data for us in this game. There are a ton of variations in the game of pool and our AI models tend to get some things wrong quite often.
We are hoping to release a beta version sometime next year, but we are a small company and there are many projects we are currently working on. As we see it, AI models in computer vision and generating the correct stats can be trusted about 50-60% of the time. It took AlphaGo a team of 30-40 of the best AI researchers in the world to develop their model and it took them two years. So be patient and plan to do many mundane tasks such as categorizing image data for thousands of images.
You can see our press release here:
http://poolstats.ai
We are hoping to release a beta version sometime next year, but we are a small company and there are many projects we are currently working on. As we see it, AI models in computer vision and generating the correct stats can be trusted about 50-60% of the time. It took AlphaGo a team of 30-40 of the best AI researchers in the world to develop their model and it took them two years. So be patient and plan to do many mundane tasks such as categorizing image data for thousands of images.
You can see our press release here:
http://poolstats.ai
Sounds like the first Chess AI
for pool it could work in 9 ball.
it could work but It might only be able to predict at most 1 or 2 shots in advance.
I would need time with the data.
The sad part is they got paid. Then there is always a research group that breaks the tech. I was in the research group.
I also read that as a way to promote science careers. Typically a project like that is over funded and not long term. Do you know if the development team stayed with the product or left after consulting was over? The remaining staff usually just maintains code.
wut? DeepMind is an essential part of everyday Google Activities now. I'm sure it consists of 100s of the leading AI experts.
Yeah, it would be interesting for sure.
I think in 9B it would be able to predict further then 1 or 2 balls but the accuracy would be affected depending on the amount of data. If a guy has 10K games it would clearly be more accurate then a guy with 100 games.
In 8B it could determine the best way to play the balls based on your skill. Shoot a different ball and it recalculates.
There used to be a lot of engineers for magnetic storage too. AI is trending now but that tech was developed ages ago. Did you come up in the engineering schools?
Great proposal, to solve it can take some planning skills, whats your plan?
Talk about it on an open forum so someone else with the skills can develop it. Then, when it doesn't work the way I want I can complain about it
Theoretically, it's possible given what we know about AI. However, it is no easy task and one person cannot do it. I don't care how much time you have on your hands. Trust me. Our vision is that it will take some new developments in machine learning to get it anywhere around 90% accurate all the time. Even then, there will be need for manual input to alter the way it behaves, for numerous test cycles before it becomes usable in the mainstream. Right now, focus on the smaller more manageable tasks such is knowing when a ball is potted, what ball was potted, the angle it was potted at and trajectory to the next ball.
Good luck!
Good luck!
Did you use transfer learning on one of the imagenet models for ball identification or build a custom network using inputs specifically tailored for pool balls?