Hello @julianStreibel
opening this issue here to collect ideas and requirements for extensions around the now published multiarmed bandits (MABs)
In the team we already talked about uniting the different variants under one MultiArmedBandit surrogate class so that one does not have to choose the right class for each target/prior type, but the surrogate does this for you. This was started in #372 but this is on hold depending further discussion
Apart from that, what other bandit classes do we envision? Do you already have some coming from business requirements?
Naively speaking, I'd like something for
- numerical targets (ie not binary)
- other priors
- refacor code so that all encoding types and discrete searchspaces are allowed
There is this nice paper (with a lot of testing around the bandits in SI) that uses bandits for chemistry, perhaps we can draw inspiration on what architectures to favor https://www.nature.com/articles/s41586-024-07021-y
Hello @julianStreibel
opening this issue here to collect ideas and requirements for extensions around the now published multiarmed bandits (MABs)
In the team we already talked about uniting the different variants under one
MultiArmedBanditsurrogate class so that one does not have to choose the right class for each target/prior type, but the surrogate does this for you. This was started in #372 but this is on hold depending further discussionApart from that, what other bandit classes do we envision? Do you already have some coming from business requirements?
Naively speaking, I'd like something for
There is this nice paper (with a lot of testing around the bandits in SI) that uses bandits for chemistry, perhaps we can draw inspiration on what architectures to favor https://www.nature.com/articles/s41586-024-07021-y