Mr 'Don't Do Evil' has patented deep Q learning,it is getting a patent on brain

>Mr 'Don't Do Evil' has patented deep Q learning,it is getting a patent on brain

Don't editorialize submission titles like this, please. The actual title is (and should be posted as) "Methods and apparatus for reinforcement learning."

Yet another attempt to own everything.

They even reference Martin Riedmiller's work from 2005 on "Neural Fitted Q Iteration" and dismiss it, saying that his paper was just an "attempt" to use an MLP for the Q function, and that because his paper talks about rebuilding a new network with each new batch of data, resulting in an inefficient learning process - ergo, their method is novel.

What utter nonsense. This completely misrepresents Riedmiller's work. They are really splitting some fine hairs here. Also, they admit that Martin Riedmiller wrote about the idea, even though he didn't use it in that paper.

And yet, they still hope to control the technology.

I am tired of FUD stories about how broad a patent is. Patent claims are organized as a series of claims with each claim narrower than the one before. Patents are always written with the first claims extremely broad and are not expected to be granted. It isn't until near the end of the claims that the meat of the patent exists.

In this case it has something to do with using multiple neural networks to train each other. I've read that is what they did with alphaGO.

The patent system is patently broken.