This project is an implementation of Population Based Training. Population Based Training is an asynchronous optimisation algorithm to optimise a population of models and their hyperparameters. The implementation is based on: https://arxiv.org/abs/1711.09846
The project consists of two parts:
- Application of Population Based Training on a toy example
- Combining the Reinforce algorithm with Population Based Training
The objective is to maximize the function
1.2 - (pow(param_0,2) + pow(param_1,2))
We assume that we can't observe this function. Instead we maximize a surrogate function
1.2 - (hyperparam_0 * pow(param_0,2) + hyperparam_1 * pow(param_1,2))
Performance of the Reinforce algorithm with Population Based Training (5 Workers):