Streaming GPR

[kjappelbaum]

Feature description

GPRs are not optimally suited for streaming applications, but there are some approximations that should allow us to drastically cut the computational cost for the GPR step.

Implementation idea

• Use code from https://github.com/secondmind-labs/trieste/blob/9847d3f4847e732e30a5b80be25d181c4ee2aebc/trieste/models/model_interfaces.py#L272 or https://github.com/thangbui/streaming_sparse_gp

[kjappelbaum]

@byooooo any opinion what is the best option?

• making trieste or something else a dependency only for this model is annoying
• adding it can be also annoying if we do not have a direct use for it (but we would still need to maintain it)
• hypothetically, one could also add ePAL sampling to trieste, but I think this is quite unfeasible as the library pursues a different objective