New neural networks coming to PA

pff silly me i didn't see PA was right there after community's in the 2014 awards. my bad.

well good on them, it wouldn't have made sence to state anything other!

Please, don't assume that everyone who voted had no clue. That's a bit rude.

there was maybe a bit of a generalization, no assumption.

I could only agree that they have no clue. I mean assasin's creed crowds? shadow of mordor? Really?

not everyone, just most.

To take an example, my high school valedictorian selection. They took the top 10 marked students and let the class vote on them. Amongst this group were the quiet super genius, couple big time school and community volunteers etc.

Who won? the tall skinny blond girl of course.

That's fair.
On the other hand, the nemesis system, which is used in Shadows of Mordor, sounds really interesting. (Though I've not played that game)

it's not ANYWHERE near as complex as neural networks.

Bla bla bla. Some technical **** only a computer wizzkid understands. Bla bla bla.

And then the AI is smarter.

Yeeey! Good job!

Next time just say that lol cause I didn't get that stuff in your blog lol.

It was written for the wizzes though

the muggles will be muggles, nothing we can do about that lol

You'll enjoy this Sorian. Deep Mind has now been taught to master 50+ classic arcade games. Is PA next?

http://www.readcube.com/articles/10...5vvfrgpGpS6fpq8Mhwn3csTrnQW_j5o3pIdS_vUhRZg==

intriguing

well so far a "classical" "AI" approach that just flat out is programmed to do the best known thing will easily outscore any human in those games as well. So while the learning stuff is really awesome it does not yield "better" compared to rather simple "do one thing"-programs results yet.
games like PA are still quite different: Here humans dominate over classical AIs.

This strongly depends how you define "better"
The remarkable fact of the approach this paper presented wasn't that it played sometimes better than humans, but that it can learn to play so many different games just from a visual input in the first place.
A single neural net architecture successfully learned all those games without tweaking the learning parameters to the game in question. And their approach is really simple, just wire up visual output of the game to the neural net, make sure the score changes gets somehow into the learning process, and that's it. Drink tea while the neural net is learning how to play the game.
This is more or less a rudimentary form of what we did when we played those games in our early years. Just play the game a bit and if an action had a positive outcome repeat it, otherwise pick another action and try it again.
Yes, comparing the learning rate of this net with an human learning rate is like comparing a single match to a nuclear explosion, but it is a starting point.

Sure is and I am deeply impressed. Just saying that they are still far away from wiring it up to PA and beating the build in AI, let alone a real human.
Well I guess they haven't tried.

Can the current AI cheat by selecting multiple units and control them individually? Because if they can, they could be able to micro to victory

Seeing even humans having trouble with the 3d spheres known as planets in PA, I think it may take a while until we see neuronal nets playing PA

On a side note - while speaking of 3d-spheres - there is the question if the computational complexity of the 3 dimensional space is enough to efficiently solve physical problems within higher dimensional space.

Yeah i'm not seriously suggesting Sorian use DeepMind to learn PA. Although the code used there is freely available for non-commercial purposes, it's not available for PA. Further while it is key that Demis et al demonstrate their reinforcement-learning based network can learn all these games from visual input alone, this is still based on a highly customized I/O system. They actually built a custom atari emulator to be able to compress all the possible frames of each game into about 200 input frames. The AI also has a simplified version of the ATARI controller to make responses with. So it's incredibly impressive Deep Mind can do this with only the visual representation of the game (true unsupervised learning), but it's still far too purpose-built for PA. But nonetheless, interesting for Sorian i'm sure!

It's quite possible that the opposite might be the case. Humans struggle so much because most have foreknowledge which they are trying to apply to the movement on spheres. That knowledge still translates quite well to movement on the surface of a torus, but not onto a sphere since the X coordinates converge near the poles. Even less so, if the poles are traversable without any length distortion.

Explain yourself? I think I know what you mean, but that's not really a limitation. You can build an efficient solver for problems of a certain number of dimensions (mostly by constructing smart data structures which allow sublinear lookup time), but nothing stops your from extending that base for higher dimensions by adding the additional attributes to the datapoints. If you say "Give me all persons in cube X with age > 70", then you are already dealing with 4D informations.

Spheres are troubling for a different reason, not so much for the 3D part, but for the non-Cartesian coordinate system of the surface, and the violation of the triangle inequality when you try to comprehend the surface as planar.

I really don't like that sentence "depends how you define better". In a game "better" is perfectly defined.
Anyway that stuff is really interesting. I think a good way to make general purpouse AI could start with it resolving games and abstracting some meta knowledge from that.