To implement agents that can learn through social interaction as humans do, I decided to explore how humans instruct each other using causal and counterfactual explanations. I focused on reinforcement learning benchmarks and built a webpage to interact with the gym-minigrid environment. The scenario involves a teacher and a learner. The learner has to navigate the minigrid environment following the explanations of the teacher, who has full visibility about the reward function (i.e., how the learner “ought” to behave).
Preliminaries:
The environment has been built as part of work done at Mila. The Dynamic obstacles environment has been added as part of work done at IAS in TU Darmstadt and the University of Genoa for mobile robot navigation with dynamic obstacles.
The interface runs on a small Node.js server, with Socket.io carrying the live connection between the browser and the environment, so that the teacher’s explanation reaches the learner while it is still moving.
The point of building it as a webpage rather than a script was to collect explanations from people who had never seen the environment before, which meant the whole thing had to be usable without a terminal.
