Nikitas Rontsis

Approximate PSD projections

2

## Purpose This is a PR that uses approximate projections for the PSD cone, as suggested in [this](https://arxiv.org/pdf/1912.02767.pdf) paper (in particular, section 5). ## Minimal example Let's solve SDPLIB's `gpp-124-4`...

Currently, the tests for the Pardiso Linear Solvers do not run in CI. In order to include them in CI, we need to install both the MKL and the commercial...

Schur Factorisation Reordering fails

14

When using Arnoldi on some (badly scaled) problems, `trexc!` throws `LAPACKException(1)` with the following stacktrace ``` [1] trexc!(::Int64, ::Int64, ::SubArray{Float64,2,Array{Float64,2},Tuple{UnitRange{Int64},UnitRange{Int64}},false}, ::SubArray{Float64,2,Array{Float64,2},Tuple{UnitRange{Int64},UnitRange{Int64}},false}) at /Users/nrontsis/KrylovKit.jl/src/dense/linalg.jl:415 [2] permuteschur!(::SubArray{Float64,2,Array{Float64,2},Tuple{UnitRange{Int64},UnitRange{Int64}},false}, ::SubArray{Float64,2,Array{Float64,2},Tuple{UnitRange{Int64},UnitRange{Int64}},false}, ::Array{Int64,1}) at /Users/nrontsis/KrylovKit.jl/src/dense/linalg.jl:335 [3]...

Memory allocation

4

Thanks a lot for the very nice package. Although it seems that your package is faster and more robust than `eigs` it allocates significantly more memory than `eigs`. For example,...

See [this](https://github.com/nrontsis/PILCO/issues/31#issuecomment-524596610) comment

Use Matlab's implementation to train OpenAI gym tasks

1

This will allow easier debugging and comparison between the two implementations.

Make trigonometric augmentation user friendly

2

Matlab's implementation augments the variables that represents angles to two new states that are simply their cos and sin. This is important for the performance of the algorithm and is...

Extra noise in controller's input covariance

2

I've noticed that the original implementation of PILCO adds an extra variance, equal to the noise of the underlying GPs, in the controller's input. This is done in [this](https://github.com/nrontsis/PILCO/blob/a4881fc72dc8ddb518b7a65cb30e01a8a83cb46e/tests/Matlab%20Code/propagate.m#L53) line....

Initial implementation of sequential batch optimization

2

This is a first attempt to partially address #71 Minimal example (modified version of [this](https://github.com/GPflow/GPflowOpt/blob/master/doc/source/notebooks/firststeps.ipynb)): ``` import numpy as np from GPflowOpt.domain import ContinuousParameter import GPflow from GPflowOpt.bo import BayesianOptimizer...

Batch Optimization

11

Hi, I am really interested in the project and, since I am working on Batch Bayesian Optimisation, I would be keen on extending GPflowOpt for it. Would you be interested...