Philosoraptor
Well-Known Member
Hi all,
I came across something today that took my eye so I thought I would share.
I came across something today that took my eye so I thought I would share.
Still not SWOS! (1 Viewer)
- Thread starter Philosoraptor
- Start date
Philosoraptor
Well-Known Member
Yes, its an evolutionary pattern of thinking in football. The more it plays, the more it learns. Would be interesting to see what it would come up with after a few months of running.
Last edited:
Philosoraptor
Well-Known Member
Its the way forward though.
Nick
Administrator
Whats the difference with football manager or the computer on FIFA?
Philosoraptor
Well-Known Member
Loads, its like comparing a server with a computer.
Philosoraptor
Well-Known Member
The programme maybe be basic in its shell but it shows the sort of framework you need to crack this idea.
Houchens Head
Fairly well known member from Malvern
I still prefer my old tennis game though!
Philosoraptor
Well-Known Member
Sorry, football in general.
Nick
Administrator
Philosoraptor
Well-Known Member
Two great videos on this.
Philosoraptor
Well-Known Member
It's only a matter time before something is going to come out which will be the football version of Chessbase or Ubisoft's Chessmaster.
Nick
Administrator
What will that do though?
Still don't get the link, chess pieces can only do certain things and move into different places.
Footballers are dynamic.
Philosoraptor
Well-Known Member
Basically what you're watching in the video are the basic ideas which are at the core of this even though the heuristics are all different.
Houdini Chess Engine
Houdini Chess Engine
Last edited:
Philosoraptor
Well-Known Member
Yes, it gives the starting point for performance analysis which can also be worked into any future work.
Nick
Administrator
Doesn't that already exist though with GPS, pass completion, run completion, tackle completion etc?
Ian1779
Well-Known Member
I was listening on the radio talking about football stats like pass completion, they were saying the next stage was to weight the quality of the passes/tackles/runs etc. John Stones has the highest pass success rate in the PL - but quantify that against someone like Alexis Sanchez who doesn't even rank in the top 200.
Philosoraptor
Well-Known Member
It just gives individual assessments. It's not the greatest thing to base any analysis of team sports off.
Have a look at this. What the engine is doing is going through the game and comparing it to all the herustics it has in its engine and seeing the deviation from them within the game. It's not comparing that particular game with any other game even though the opening book can do this.
What you're looking at is the front end of what's already been described in the thread with all the bells and whistles added.
Last edited:
Skybluefaz
Well-Known Member
Surely it would only be of use if robots were playing football not humans.
Philosoraptor
Well-Known Member
I just don't think they have a way of doing positional analysis yet. If they did then surely we would of seen this already.
Imagine having something like this running in the background comparing it to the positions of players in a team, and the decisions these players are making, if it was set up with the exact heuristics the manager wanted the team to play with.
Last edited:
Nick
Administrator
Doesn't gps do that? Heat maps?
Philosoraptor
Well-Known Member
I would say these are just rough estimates. The thing to do would be to tap directly into the heuristics of the game for your answers.
Nick
Administrator
Surely that's ok with chess pieces, but not with humans on a football pitch?
Philosoraptor
Well-Known Member
Lol, you asking question faster then I can form answers.
It shouldn't matter in all honesty. They are both points on a playing field that can be measured be it they act in different ways.
Nick
Administrator
Yeah but gps can measure points, distance etc
Philosoraptor
Well-Known Member
individually, there is no way of measuring the heuristics in the game and the choices of what are made. I guess what you are looking for is a programme to go on top of the tracking system to look at the game hereustics and the position of your team on the pitch compared with the opponents positions and see if better choices could be made within the game from any particular time.
A heat map just doesn't deliver this.
A heat map just doesn't deliver this.
Last edited:
Nick
Administrator
But.gps can do that?
Philosoraptor
Well-Known Member
Gps doesn't provide game analysis. Just positional information.
shmmeee
Well-Known Member
Phil, this is my specialist area (Machine Learning). Just to correct a few things:
- Chess doesn’t use Machine Learning. It’s solved brute force because every possible future position on the board can be calculated. So the computer just runs the numbers for every move and picks the best.
- In terms of neural nets 2010 is ancient. This uses an MLP, which is very basic and has no memory or anything higher order, just tries to map inputs to outputs.
- AFAICT it’s not a genetic/evolutionary algorithm either. The video confuses some terminology so maybe, but I’d assume it uses gradient descent for learning rather than a GE.
A better example of something that could evolve into a proper football sim would be DeepMind’s AlphaGo work, or their Atari playing stuff. This sort of high level planning and long term goal seeking would probably do better with a reinforcement learning approach than supervised learning.
If you’re interested this is a decent video on AlphaGo and another on the Atari stuff both from Google DeepMind.
And here is the latest RoboCup which is a better description of state of the art.
I’m not an RL specialist by any means, I work with ConvNets and regression mostly, so may have got some of the finer details wrong.
- Chess doesn’t use Machine Learning. It’s solved brute force because every possible future position on the board can be calculated. So the computer just runs the numbers for every move and picks the best.
- In terms of neural nets 2010 is ancient. This uses an MLP, which is very basic and has no memory or anything higher order, just tries to map inputs to outputs.
- AFAICT it’s not a genetic/evolutionary algorithm either. The video confuses some terminology so maybe, but I’d assume it uses gradient descent for learning rather than a GE.
A better example of something that could evolve into a proper football sim would be DeepMind’s AlphaGo work, or their Atari playing stuff. This sort of high level planning and long term goal seeking would probably do better with a reinforcement learning approach than supervised learning.
If you’re interested this is a decent video on AlphaGo and another on the Atari stuff both from Google DeepMind.
And here is the latest RoboCup which is a better description of state of the art.
I’m not an RL specialist by any means, I work with ConvNets and regression mostly, so may have got some of the finer details wrong.
Philosoraptor
Well-Known Member
This is excellent stuff.
I remember following the Go games online when they took place.
The engines do use variations around a minimax algorithm with alpha beta pruning which is basically a brute force approach.
This should be a lot easier to explain.
Let say an engine is set up with opening theory and an endgame databases already in its system, you can use a football analogy of attacking and defending. It would be the middle game or midfield where the brute force approach takes place. Surely this can be better analysed then it is at the moment.
I think it is important to point out that chess engines aren't looking at every single move that could possibly be played in a course of a game. They are doing a depth search of a position to give the highest value return from the algorithm used. This higher value doesn't necessarily always mean piece count but a better positional advantage.
This is what beats Human opposition.
This is a great video on how far positional analysis has come.
Game Over Kasparov and the Machine - Documentary Mania
Last edited:
Make someone rich? :$
Jeez you know your stuff young man.
Chris Robinson/Robertson will/should be all over this. B-)
shmmeee
Well-Known Member
Fair points regarding brute force I massively oversimplified. But the point is they don’t really “learn” so much as play the numbers they have. DeepMinds latest Go algorithm AlphaZero is given nothing but the game rules and plays itself, it won something like 100 games in a row against the original.
It’s true it does use a similar method, but the clever bit is how it prunes the search tree. I think something like the StarCraft AI they’re working on now would be good for football as well.
I’m just a DeepMind fanboy, love that a British company is leading the world in AI. Also the founder did the AI for Theme Park and that was amazing.