Inferring models for collaboration using Turing Learning

This project explores Turing Learning [1], a method that infers the behaviour of a given system without needing pre-defined performance metrics. Turing Learning enables robots to model themselves, other robots, and the environment.

Turing Learning refers to a family of machine learning algorithms that generate models and discriminators in a competitive setting [1]. A training agent, T, provides genuine data samples, while a model agent, M, provides counterfeit data samples. A discriminator agent, D, labels the data samples as either genuine or counterfeit. Agents M and D are being optimised: D is rewarded for labelling data samples correctly, whereas M is rewarded for misleading D (to label its data samples as genuine). Turing Learning is inspired by the Turing test. It is a generalisation of generative adversarial networks (GAN) [2], and can be used to actively infer the behaviour of agents through interrogation [3].

The project is expected to improve a robot’s ability to model itself or others and to leverage such models in collaborative settings. Collaborative tasks involving swarms of robots or mixed human-robot teams could be considered. Experimental validation with real robots is encouraged.

The applicant would join a vibrant group with excellent facilities (http://naturalrobotics.group.shef.ac.uk).