thank you!!

glad it is useful 🤗

@@mlandaiacademy Trust me its too good and i will be completing the remaining two tomorrow. I really hope you make videos on GAN, CNN, RNN too.

@@SuvradipDasPhotographyOfficial Thank you. We will try if time allows for sure :)

glad you find it useful.

Glad it helped!

We are planning to put the paper out soon we are just writing it - stay tuned.

Never mind. He mentions paper is coming, and they are working on how to share the code.

He is me btw hehe

@@mlandaiacademy Then let me thank you for a very interesting video! I am working on my Phd. project proposal right now, and I am looking to work on embodiment of robots and navigation in Agriculture. This domain is exploding right now!

We should deffo have a chat write me on [email protected] if interested 😀

Yes KZbin has two qualities sd and hd. SD processing is done and HD processing typically takes more time. In couple of hours if you refresh it should be hd quality 😀 it’s KZbin’s processing time 😀😀

Thanks a lot Tarek! You are the best. I am glad you find them useful ! Appreciate the kind words as well 🙏🤗

Thank you so much!! Glad you like it 😀

Thank you so much !!!

Think about what would happen if we do multiple Alpha by a vector and take the l0 norm of that. Then compare that with what the axioms of norms should give us for a the norm of the vector alpha x vs the norm of vector of x

Thanks a lot glad you find them useful !

Of course :)

You are very welcome !!

Thank you so much!!

متشكر اخي محمد ☺️

Glad you enjoyed it!

This is an interesting way to see it. The way I see it is like it is trying to design denser rewards from a sparse signal in a way. Based on fitness it is trying to make the problem easier to solve by giving me easier rewards at each stage. It’s something similar to what u mention yea

No you aren’t. It is a short teaser to a much longer paper. It is more saying go read that paper it is awesome.

I like it too yes 😃

Thanks you so much!!

Glad you find it useful!!

You are very welcome !!

computer no good keeping context

Glad it was helpful!

First, I would make a distinction between the dimensions of the problem and the values each dimension could take. In terms of dimensions, you are right that standard BO could suffer - maybe up to 80 dims I'd say if you try hard. However, there are high-d BO algorithms devised to takle higher-d, some include: TuRBO (arxiv.org/pdf/1910.01739.pdf), latent space Bayesian Optimisation (arxiv.org/abs/2201.11872) and decomposition techniques (arxiv.org/abs/2301.12844). In terms of the second quesiton, that relates to the combinatorial nature of the problem for which combinatorial Bayesian optimisation solvers have been developed. The trick is in the kernel and how you maximise your acquisitions to better explore the space. If interested, we had a new library at NeurIPS that solve those problems and gives you a SOTA algorithms. Please find the paper here: arxiv.org/pdf/2306.09803.pdf, the code here: github.com/huawei-noah/HEBO/tree/master/MCBO, and a blog about the results here: medium.com/@haitham.bouammar71/introducing-mcbo-a-modular-framework-for-comprehensive-benchmarking-and-evaluation-in-5baacab71fc6. Hope this helps :)

It is not better per-say. It is more tractable though. Think about it, p(tau) = p(s_0) \prod_t p(s_t+|s_t, a_t) pi_theta(a_t|s_t). If you want to compute the grad_theta of this it is hard for many reasoning the product, not knowing p(s_t+1|s_t,a_t) and not ness. leading to a tractable algorithm which you can use to sample to get the Monte Carlo of the expectation. The log helps us in those. When you take the log the prod becomes a sum and grad wrt to p(s_t|s_t, a_t) vanishes since it does not depend on theta. Then you are left with grad over pi_theta. Hope this helps!