REAL WORLD vs ABSTRACT VIEW of an AI
— What do you see and what does an AI see?
By Mathias Paul Babin, PhD candidate, Western University
This app is in development.
The focus is on teaching not only the fundamentals of reinforcement learning, but also on how machines ‘see’ the world.
The goal of this app is to control a robot’s actions such that the user maximizes their score/reward. Initially the environment is represented with basic geometric shapes and objects to show how a set of sensors on an actual machine would see the world, that is, an abstracted view of it. Next, we have the students repeat the exercise but in a world rendered with grass, dirt, and rocks to provide real world context to the task being performed.
As for the task’s reward signal itself, 2 points are given if a tree is planted in lush fertile grass, 1 in less fertile dirt, and 0 if in a patch of dirt and rocks. With context it is clear that students should be planting in grass but when this world is abstracted away from them and they are only provided with arbitrary colours to represent terrain types, it becomes an exercise in exploring the environment in order to learn what act