They characterize the convergence behaviour using the "Committal Rate" of the algorithm, quantifying how aggresive the algorithm's update rule is.

Methods benefitting from the geometry (e.g., Natural PG) behave much better than standard PG (under softmax policy), if they have access to the exact policy gradient (or NPG). If the direction of improvement is estimated on-policy with noise, PG >> NPG.

- The Chinese room argument is apparently clever, but it has never been convincing to me.

- Some stories about him:
www.facebook.com/bernard.w.ko...

1) The exponential growth of the field implies an exponential growth of recent papers, so even uniform sampling of papers means reading more of the recent ones.
2) With a limited time to read, there is a FOMO-type of incentive to focus on the new. The sampling distribution is not uniform.

But its output was inspirational!

Surprisingly, it didn't do a good job. Maybe it needed more hand-helding and guidance. My prompt was something like this:
This is my Research Proposal, these are the Evaluation Criteria, these are suggestions of what should go in each section, and this is my actual CV. Come up with something good.

What is expected is not clear, there is no samples to see, there is a 5-page limit (if you write in French, you get an extra page!), but I suppose this invention helps everyone keep busy and the beareaucrats continue to have their jobs.

I don't know what exactly he was thinking of (minimax results? adaptivity? continual learning?), but I have been thinking of a version of this thought in the past 2-3 years, 50 something years after him!

... I propose to study the synthesis of brain models by the parallel development of a series of matched (theoretical) environments and corresponding brain models that adapt to them."