I am trying to implement a DNN to optimize a set of 7 parameters that are used in a robot swarm simulator on the ARGoS platform. the program is a compiled C++ executable that reads the parameters from an XML file.
Previous experiments to optimize these parameters used GA techniques to find optimal parameter settings for specified distributions, but this technique may not be finding global maximums. We would like to try using a DNN with automated hyper-parameter tuning, specifically Vizier (HyperTune on the Google Cloud Platform) but I am having some trouble designing how this problem is solved with tools like TensorFlow. Using a GA to select for an optimized parameter set is straightforward, but how to use a DNN on TensorFlow, and take advantage of Vizier/HyperTune to tune this DNN?
I have installed the sim in a docker container that can launch the sim with a parameter set via command-line arguments, and can return the resulting score of the trial, but how I can connect this into a DNN program to optimize the input parameters by maximizing the resulting score is unclear.
I am a novice ML programmer so please pardon erroneous assumptions or methods here. Any help is greatly appreciated!
note the reason I am looking to use TF would be to take advantage of the native GCP hyper-parameter tuning tools, which appear to be tailored for TF.