Turing Learning: Advances and Applications

Turing Learning is the family of algorithms where models and discriminators are generated in a competitive setting. This thesis concerns the coevolutionary framework of Turing Learning and investigates the advances for improved model accuracy and the applications in robotic systems.

Advances proposed in this thesis are as follows: an interactive approach to enable the discriminator to genuinely influence the data sampling process; a hybrid formulation to combine the benefits of the interactive discriminator in improving model accuracy and the advantages of the passive discriminator for reducing training cost; an exclusiveness reward mechanism to promote candidates with the exclusive performance during the coevolutionary process.

Applications presented in this thesis are as follows: an approach for a mobile robotic agent to automatically infer its sensor configuration; an approach for the robot agent to automatically calibrate its sensor reading; a novel approach to infer swarm behaviours from their effects on the environment.

The interactive approach has been validated in the inference of sensor configuration and calibration model, leading to the self-modelling/self-discovery process of robotic agents. Results suggest an improved model accuracy with the interactive approach in both cases, compared with the passive approach. The hybrid formulation and the exclusiveness reward mechanism have been demonstrated in the inference of the calibration model. Results show that almost half of the training cost can be reduced without a decrease in model accuracy by applying the hybrid formulation. The novel reward mechanism can accelerate the convergence without a decrease in model accuracy. The indirect way of inferring swarm behaviours requires a small amount of training and reveals novel behavioural controllers for individual robots.