This repository contains prototype implementations for my work on learning optimised task-mappings via reinforcement learning. Here, I aim to leverage Graph Convolutional Networks to learn mapping strategies that incorporate awareness of the placements and features of the other related tasks that scheduled on the same hardware. I therefore use code from Kipf and Welling's GCN repository here.

The repository is split into two implementation folers. The src_simple_mlp folder provides a simple implementation that uses the REINFORCE (and its baseline adaptation) RL algorithm to train a Multi-Layer Perceptron. The src_ppo_gcn folder provides a more complex implementation that uses Proximal Policy Optimisation to train a GCN model in an Actor-Critic setting, and also incorporates Random Network Distillation to encourage exploration.

The prototype code as well as this README represents work-in-progress.