Re: NN for Pinball Physics

I think a back-propagation nn can do the job.
A quick way to train it is to collect data from a real game: simply
execute/train the net on the input/output generated by a player playing
pinball, and then see if the network has learned to play.
But I think the only relevant situations are those when the ball is
close to the flippers, because it is difficult to make the network
understand and generalize the irregular movement of the ball bouncing
on various objects.
May be you can have a look at a similar program I have written, a
neural network that learns to play pong. It is only an experiment so it
is not documented, it's written in Python and C and uses my "Neural Net
Framework".
http://alessandropresta.altervista.org/neuralpong.tar

Bartosz Bien ha scritto:

> Hello,
>
> We have just developed a Pinball game and I came up with an idea that
> it would be interesting to develop an autoplayer AI for testing purposes.
>
> Given the location and velocity vectors, the AI player would have to
> press or release either left or right flipper handle(s). Although all
> objects in the scene are 3D, th physics is simulated in 2D, so both vectors
> are two-dimensional. Therefore, the problem has four real inputs (x, y,
> vx, vy) and two binary outputs (left/right flipper pressed=1, released=0).
> Each table should utilize its own network, because each has a different
> distribution of action objects.
>
> ***** May you have any suggestions on how to train the network best? *****
>
> You are welcome to take a look at our webpage (below) for further
> information on the pinball project (demo is also available from the
> distributor).
>
> --
> Best regards from Poland (there's a first snow this season!),
> BB
>
> http://spin.neostrada.pl

.