the neural network deep learning for computer player?
Moderators: hoijui, Moderators
the neural network deep learning for computer player?
do you think it is now possible to create an AI called deep learning for the player of the computer or it will require too many resources? And if so about what?
- Forboding Angel
- Evolution RTS Developer
- Posts: 14673
- Joined: 17 Nov 2005, 02:43
Re: the neural network deep learning for computer player?
Search the forum for neural network. It has been discussed at length in the past. Perhaps those discussions might shed a little more light.
Re: the neural network deep learning for computer player?
There's only one way to find out. Fork one of the Java/Python/C++ AIs and do stuff with Tensorflow or one of the other frameworks
Re: the neural network deep learning for computer player?
To train a NN AI on a specific Spring game would be possible.
For this, the conventional replays have to be taken, and the data filtered into something a NN can percive and derive conclusions on.
This means, there has to be a Orders given to Unit Dataset.
There has to be a derived Dataset of Unit of Type(x) Dataset showing the execution of the order.
There has to be a derived Dataset of Health and Unitstate.
There has to be a derived Dataset of Projectiles fired.
There has to be a reward function, which deduces from health loss and area claimed which movements where actually positive.
Remeber that the Neural Net might not understand "hidden" corelations. It does not understand what a medic is- it will only correlate that units near a certain type experience health gains.
Now the NN is trained against these datasets derived from replays. For this to be a general NN regarding Maps, it will also have to know the metallmap, the heightmap and needs replays on various maps.
If you are happy with a AI that "only" plays Goddelich- this will result in such a AI. It may still utterly fail on a not seen before map, with unknown conditions- or with not known units.
If you want to go forth, you can do reinforced learning, in which the NN plays against itself and improves based upon the data of those games. This is the approach that Alpha Go used.
Would love to see a generic Spring NN, that - knowing the Unitdefs, is able to play with everything thrown at it - but beware that this will suck up a ton of GPU calculation ressources and LUA magic is still transparent to it.
PS: My epxerience on NN is very limited. I tryied to train a chat-support bot, but honestly didnt have the machinery to run the training 24/7 for months.
https://github.com/PicassoCT/neuralconvo
For this, the conventional replays have to be taken, and the data filtered into something a NN can percive and derive conclusions on.
This means, there has to be a Orders given to Unit Dataset.
There has to be a derived Dataset of Unit of Type(x) Dataset showing the execution of the order.
There has to be a derived Dataset of Health and Unitstate.
There has to be a derived Dataset of Projectiles fired.
There has to be a reward function, which deduces from health loss and area claimed which movements where actually positive.
Remeber that the Neural Net might not understand "hidden" corelations. It does not understand what a medic is- it will only correlate that units near a certain type experience health gains.
Now the NN is trained against these datasets derived from replays. For this to be a general NN regarding Maps, it will also have to know the metallmap, the heightmap and needs replays on various maps.
If you are happy with a AI that "only" plays Goddelich- this will result in such a AI. It may still utterly fail on a not seen before map, with unknown conditions- or with not known units.
If you want to go forth, you can do reinforced learning, in which the NN plays against itself and improves based upon the data of those games. This is the approach that Alpha Go used.
Would love to see a generic Spring NN, that - knowing the Unitdefs, is able to play with everything thrown at it - but beware that this will suck up a ton of GPU calculation ressources and LUA magic is still transparent to it.
PS: My epxerience on NN is very limited. I tryied to train a chat-support bot, but honestly didnt have the machinery to run the training 24/7 for months.
https://github.com/PicassoCT/neuralconvo