“we encourage approaches that adaptively explore the parameter space, guided by simulation feedback” I doubt this is possible why within 1hr, you can’t solve an ill-posed problem by calling simulation like 6times, it might land on the wrong source configuration,
Where do you see the 1h limit?
Also, I assume we can solve multiple problems in parallel. My simulation budget is far larger than 6 calls per sample.
The last paragraph of The Evaluation section says:
”…To encourage you to find efficient solutions, each final submission must run on a similar number of samples as the test dataset in under 1 hour on a G4dn.2xlarge machine on AWS.”
I haven’t found the “~6 calls per sample” though.
If parallel execution is allowed, then the maximum is `120 samples per call.
Let’s do the math quickly:
80 samples → 1 hour, right?
That gives 45 seconds per sample. Now, the simulation itself takes 3 seconds to run, so we can make 15 calls within that time. I said six earlier because I was trying to be extra conservative — idealism doesn’t exist in the real world, aside from ASML 
.
Anyways, if we run in parallel, that becomes 8 × 15, which brings everything to a maximum of 120 calls per sample.
I wanted to write a custom CUDA kernel for the simulation, but guess what , they said we should use it as it is 
,
Honestly I was shocked to see 1.22xx, max score is 1.30, so now subtracting 0.30 means 0.08 error margin which is so impressive without large calls, there is still time though for experimenting.., without the limit of 1hr my most accurate solution would have taken like ~2weeks for inference (I didnt bother submitting it)