Sorry for the big delay in continuing this project, I was away on holiday for a large part of summer and have been dealing with NeurIPS submissions / reviews and writing my thesis since coming back. But I have finally found time again for finishing the multi-fidelity strategies!
In this pull request you can find an initial attempt at incorporating transfer learning BO using MultiTask GPs. I think I almost got it working, with very few changes to the code. However, there are a few avenues we can take and I was wondering which one you would prefer. I included a jupyter notebook, transfer_learning_bo.ipynb with this pull request detailing what I want to achieve and the changes that I think may be required in the code.
Once the transfer learning BO is working, multi-fidelity is a simple step-away, as all we would need to add is a way of choosing the fidelity having already chosen a candidate with transfer learning BO.
Hi @jduerholt!
Sorry for the big delay in continuing this project, I was away on holiday for a large part of summer and have been dealing with NeurIPS submissions / reviews and writing my thesis since coming back. But I have finally found time again for finishing the multi-fidelity strategies!
In this pull request you can find an initial attempt at incorporating transfer learning BO using MultiTask GPs. I think I almost got it working, with very few changes to the code. However, there are a few avenues we can take and I was wondering which one you would prefer. I included a jupyter notebook,
transfer_learning_bo.ipynb
with this pull request detailing what I want to achieve and the changes that I think may be required in the code.Once the transfer learning BO is working, multi-fidelity is a simple step-away, as all we would need to add is a way of choosing the fidelity having already chosen a candidate with transfer learning BO.
Best, Jose