idavidrein

Hey HN, I wrote an intro to reinforcement learning (RL) tutorial that tries to balance technical depth and high-level breadth. I hope this can be a useful intro to the field for people who are interested in RL, but don't want to read tons of papers that assume lots of experience with the basic ideas. I've also tried to position this to outline the methods and style of the field, so you can use your excitement/interest from this post to help gauge whether or not you want to explore RL more deeply. At the end of the post I've curated and described some of the best online (and free) resources you can explore if you want to learn more.

I hope people find this useful, and I'd also appreciate feedback and constructive criticism on the writing style, content, organization, etc.

Thanks!

You're not thinking broadly enough about the utility maximization framework. What's wrong with just including bread baking, hanging out with friends, etc. as things you find valuable, and then try and maximize those jointly with the value you find in productivity and working?

that's a fair point, but vision is such a high-dimensional and dynamic input that my feeling is the last 20% of super-fast autofocus, auto color balancing, high dynamic range, removing motion-blur, etc. will be very difficult to improve upon. it seems that just increasing brightness/fov by making better waveguides or MEMS will be much easier. that being said, magic leap's lack of success doesn't bode well.

yeah, it's possible that the stuff doesn't look very good, but my guess (maybe my hope?) is that it's too cluttering or through careful analysis could reveal IP about their predictive algos

given the rate of improvement in camera tech, I'd say we are very far away from passthrough to be usable

that's a fair point, but it seems reasonable to me that they would separate the sensory input and the predictive/higher-level aspects of their modeling. For example, we know for a fact that they must be doing tons of prediction for both cars and people, so I think it's likely that different models (not sensory) have the info that a person is probably still there.

well the ideal cheap ride-sharing AV world would have much more spread out cities, no?

I agree with a significant amount of your point, but with regard to object permanence, I would guess that they have prediction algorithms that don't only rely on the current-time perception, so if something blips out of sight for a second the system will still infer/predict it's existence (for a time - obviously if something is hidden for a long time it won't continue to not trust perception).