Nikitas Rontsis

I think I agree with @NicolayP. You might want to use sparse gaussian processes, which are [already implemented in PILCO](https://github.com/nrontsis/PILCO/blob/master/pilco/models/smgpr.py). Let me know if this helps with your problem.

@kyr-pol what do you think?

Calling python from Matlab is [very limited](http://uk.mathworks.com/help/matlab/matlab_external/limitations-to-python-support.html) (see also [here](https://uk.mathworks.com/matlabcentral/answers/314283-how-can-i-use-tensorflow-library-from-matlab-environment)), so the best way would be for Python to make multiple calls to Matlab, in the same way we do...

Sure, PRs are welcome! While coding, keep in mind that the purpose of this package is to provide researchers/experts of the area of gaussian process dynamical systems with a basic...

Let's do this, at least for the sake for compatibility.

I'll try to change this so as to use `gplfow`'s saver

The reward is calculated [here](https://github.com/nrontsis/PILCO/blob/master/pilco/models/pilco.py#L126-L136), so perhaps you could modify it to add what you need?

Thanks @nknudde @javdrher for the quick replies! I believe that both approaches can be easily incorporated in the current framework. - For *sequential optimisation* (e.g. [q-EI CL](https://link.springer.com/chapter/10.1007%2F978-3-642-10701-6_6), [BUBC](http://jmlr.org/papers/v15/desautels14a.html)) - Usually...

@javdrher Thanks, I think I agree to all of your points. Sorry I didn't notice `batch_support` before. I will proceed in implementing the batch ''sequential'' case in a context manager...

I believe that [this](https://github.com/GPflow/GPflowOpt/blob/batch_support/GPflowOpt/acquisition/acquisition.py/#L207) line should change to: ``` acq = self.build_acquisition(*tf.split(Xcand, num_or_size_splits=self.batch_size, axis=1)) ``` But, more generally, why do we require the input `X` to be [at least 2d](https://github.com/GPflow/GPflowOpt/blob/batch_support/GPflowOpt/bo.py#L188)?...