Supervised learning formulation
Hello Proxima Team!
Is my understanding correct that you are looking to build an ai-based tool which (1) consumes desired metrics like aspect ratio, elongation, magnetic mirror ratio, etc. (from ArXiv paper, Table 2), and (2) produces an optimized plasma boundary?
As if you ran a set of tools to produce the candidate boundaries, ran Oracle, and selected the 'best' of the resulting optimized boundaries.
Thank you!
Sasha
Or you are looking to approximate an Oracle and subsequent evaluation?
Essentially, is it a mapping boundary-to-metrics, metrics-to-boundary, or something in between?
Hi @orado . So far it's really an optimization benchmark. We simply want a boundary (for problems 1 and 2) and a set of boundaries (problem 3) that optimize the objective(s) while being feasible (meeting all constraints).
Learning a mapping between boundary to metrics could facilitate / accelerate this optimization. Whether the current available data is enough to achieve this / or you need to further sample data points to effectively do surrogate-based optimization is still an open question.