search

nrontsis / PILCO

Bayesian Reinforcement Learning in Tensorflow
MIT License
311 stars 84 forks source link

SMGPR : the induced points are different for each model #42

Open patxikuku opened 4 years ago

patxikuku commented 4 years ago

In the implementation of calculate_factorization of the class SMGPR, Z_0 (induced points of models[0]) is used for all the SGPR models but since the models are optimized separately, all the sets Z_i are different. In the original PILCO implementation all the models share the same induced points. Could this affect the performances of SMGPR ?

I saw that GPFlow seems to be able to handle shared induced inputs. https://gpflow.readthedocs.io/en/master/notebooks/advanced/multioutput.html

patxikuku commented 4 years ago

GPflow multioutputs is not available for the moment for GPR and SGPR. https://github.com/GPflow/GPflow/issues/1209