RPi powered AI beats human pilot in dogfight

[jbeale]

RASPBERRY PI-POWERED AI BEATS HUMAN PILOT IN DOGFIGHT
http://www.newsweek.com/artificial-inte ... -ai-475291

"When the Raspberry Pi computer was first launched in 2012 to promote the teaching of computer science, its creators probably didn’t imagine the $35 device would one day take on a professional fighter pilot in a dogfight—and win.

But that is exactly what a doctoral graduate at the University of Cincinnati set out to do when he built a Pi-powered artificial intelligence pilot. The AI, dubbed ALPHA, went up against retired United States Air Force Colonel Gene Lee in a series of simulated battles, beating Lee in every single engagement."

[W. H. Heydt]

Slate has also picked up the story.

http://www.slate.com/blogs/future_tense ... gence.html

Edit to add...
The Register has picked up the story: http://www.theregister.co.uk/2016/06/28 ... dogfights/

They've added some information...the human was given advantages in the form of more and longer range missiles and simulated AWACS intel to help him. He still lost.

USAF is going to have to add "Powered by Raspberry Pi" to it's fighter jets pretty soon... (Maybe that's where all the Model A+ Pis have been going?)

Edit to add some more....
The BBC weighs in: http://www.bbc.com/news/technology-36650848

[fruit-uk]

Suppose we are going to see a deluge of threads asking how to replicate this

Another step towards the end of us all!

[NicoVanDerDussen]

Interesting!

Any info on which Model Raspi was used?

[alynsparkes]

Interesting!

Any info on which Model Raspi was used?

Pretty sure it's NOT running on a Pi. According to the actual published PDF (http://www.omicsgroup.org/journals/gene ... 000144.pdf):

"Again, ALPHA can currently operate alongside AFSIM on a single 3.2 GHz core of a CPU."
Which Pi has a multi-core 3.2GHz chip?

[jamesh]

Interesting!

Any info on which Model Raspi was used?

Pretty sure it's NOT running on a Pi. According to the actual published PDF (http://www.omicsgroup.org/journals/gene ... 000144.pdf):

"Again, ALPHA can currently operate alongside AFSIM on a single 3.2 GHz core of a CPU."
Which Pi has a multi-core 3.2GHz chip?

The stories are a bit inconsistent, but I think the code IS running on the Pi, but the AI itself was programmed up on a larger machine prior to putting it on the Pi.

[Gavinmc42]

So a better floor cleaner robot should be easy now?

[jbeale]

On that last page, that paper also says: "Thus, ALPHA can currently operate in the domain of microseconds; one of the side-benefits of fuzzy control. This is utilizing low-budget, consumer grade products."

When you have the source code, you can compile and run software on any machine with enough resources for your program, and I presume they chose the Pi for one set of tests, to demonstrate how it could run effectively even on small/cheap hardware for the publicity impact. Which apparently worked given the discussion here .

[W. H. Heydt]

The stories are a bit inconsistent, but I think the code IS running on the Pi, but the AI itself was programmed up on a larger machine prior to putting it on the Pi.

I would suspect it of being the other way around...develop on a Pi and then port to a bigger/faster machine to run.

[jamesh]

The stories are a bit inconsistent, but I think the code IS running on the Pi, but the AI itself was programmed up on a larger machine prior to putting it on the Pi.

I would suspect it of being the other way around...develop on a Pi and then port to a bigger/faster machine to run.

My indirect experience with neural nets was that we were unable to program the net with currently available equipment - more oomph was needed. Not sure that applicable here though.

[dom]

My indirect experience with neural nets was that we were unable to program the net with currently available equipment - more oomph was needed. Not sure that applicable here though.

Correct. Training a neural net is expensive.
Running a neural net is cheap (just a weighted sum of inputs).
Generating the inputs may or may not be expensive. Often you analyse an image to produce some "features" that are the inputs for the neural net. That analysis may be expensive.

[jbeale]

Seems like neural networks in image-recognition research are certainly using a lot of hardware, based on quotes like:

https://research.googleblog.com/2016/04 ... -with.html
"Using the distributed trainer, we trained the Inception network to 78% accuracy in less than 65 hours using 100 GPUs."

[Gavinmc42]

Raspberry Pi Zero super cluster for training/learning the Fuzzy expressions?
100GPU's = 50 Zero's = $250
Then run the Fuzzy stuff on one Zero = $5 against a pilot with $$$$$$$$ of training.
What's the current cost of F22 and F35?
Replace drone operators with Pi's?
MIB going to show up on RPF's doorstep with export restrictions?

Interesting ideas here for computer gaming companies.