Youve seen backprop defeat people with their own tactics. done.
Youve seen creatures learning how to walk, even like a person. done.
both cool things, but what else can you do?
TRIPPY NET GAMES
what people dont realize, is nets can control a video screen too, and thats what i never saw, i did never know.
maybe because they require the horsepower only a modern video card can give you... but back in the old days,
you could have done it on a 16x16 binary screen, instead... and that has possibly some joy attached.
so basicly, you can record a game with them, seems kinda pointless... cause why didnt you just play the game?
and its more pointless when i tell you theres only a finite amount of different screens you can store...
as if you simply recorded it with a screen recorder and each screen is a state... the naive implementation
is a single cell per frame but ill explain later how you can increase your state size theoretically.
about all you got over a screen record, is a method of compression, and screen matching powers... but this
is the fun bit about it.
but now what can you do?
a) you can control the game inside the record, by matching screens to the fresh controls, so you can play it
out of the record. so thats the beginning.
b) it can be ai... the computer can pretend hes you if he learns off your play, and he biases the screen
to be the closest matched state, and comes up with your controls.
c) you can even hack the screen, and itll bring up the most similar screen in the record, so you can hand
around a little man, and maybe a bird, and then itll start playing where you plotted it in.
d) noise can be introduced to some parts of the input to sway the game randomly around the place, and make
it more exciting.
e) you can search for different areas of the game by giving it labels, in a special label area it learnt
whilst it was recording
f) with a bigger network with decent enough res and brightness you can teach it real video, and then if you
label the frames with controls and states, youll be able to play a game that looks real.
INCREASING STATE SIZE
well, the first thing you can do, is make your cell count a 4096x4096 texture, and thatll boost you up
to 16 million frames right then and there. itll take you ages to fill that. To reduce synapses you can
introducea a small hierarchy (making it a deep net) thatll group the screen more gradually, then if there
were 20 active cells in the second to last region, the whole states are now only 20 synapses each, for example.
theres a heaps more complex way to do it, and its by using a distributed representation to restore whole
states. its confusing, because the cells have to work as a team to some level of confidence produce
the parts that make a whole image, instead of garbage, which is what youd get otherwise.
But then you could have possibly a billion states, and youd never ever fill it!
And not to forget, if you store animations in your cells, they have to predict the animation to get them
to activate straight away when you need them, then youll naturally develop alot more states also.
This will have to be used in combination with the distributed representations, and its very tricky, getting
it right.
CONCLUSION
the sad thing is, its not intelligence, and it must be 100% supervised everything it "knows" cause it
has to record it. youll have to show it everything explicitly, so even if you had the states, making
the game that looked real would be impossible cause there would never be enough distinct video, for it
to fill it states up much at all.
The "real sim" would be a robotics venture, where you recorded the steering wheel with a camera on the
bonnet, and they would drive and drive and drive with the video autolabelled, but it would still take
god knows how many years to finish an explicit state set which could actually be interactive... if the
net was made professionally, distributed and predictive they probably never even hit capacity.
