DuelingDQN Agent with SNGNN reward

Introduction

Following the previous posts, now we are training the Dueling DQN agent but the only difference from the previous experiment is that the reward comprises of both the hand-engineered reward function and the output of SNGNN network.

About SNGNN

SNGNN is a Graph Neural Network to model adherence to social-navigation conventions for robots. Given a particular scenario composed of a room with any number of walls, objects and people (who can be interacting with each other) the network provides a social adherence ratio from 0 to 1. This information can be used to plan paths for human-aware navigation. SNGNN can be thought of a learnt reward function and hence can be used for the current RL task. Please refer to this page for more details: SNGNN

Reward Curve

The training has not yet completed but this is how the reward curve looks like

Performance

https://user-images.githubusercontent.com/39992294/128668213-b99d43b8-c448-42bd-a44f-3550661c52c6.mp4

Future work

We will be now trying to evaluate the performance of the agent with and without SNGNN reward func while training the Reinforcement Learning algorithm.