Na predavanje Piškot prihaja Timothy Wiley z Univerze New South Wales iz Sydneyja. Ukvarja se z umetno inteligenco v robotiki, že leta pa sodeluje tudi z nekaterimi sodelavci Laboratorija za umetno inteligenco. Dobimo se v petek, 16. februarja 2018, ob 11. uri v predavalnici 22. Predavanje bo v angleškem jeziku. 

Data Efficient Learning of Robot Behaviours

Povzetek:

Autonomous robots execute complex behaviours to operate and perform tasks in real-world environments. Machine learning has been employed to acquire such behaviours, usually by trial-and-error learning, however, this process often requires a large number of iterations.

If an online learning process must be performed, that is, learning on board the robot as it operates, simplistic approaches to trial-and-error learning are infeasible. The robot will break down long before any significant progress is made. Instead, a multi-stage framework improves the feasibility and efficiency of learning robot behaviours is through a domain independent approach. The framework builds a qualitative model of the robot from behavioural traces that are collected as the robot operates. The model trades accuracy for domain independence in elevating the skill acquisition problem into the high-level symbolic realm. A forward-chaining planner subsequently finds a sequence of actions that the robot must perform to carry out its assigned task. A second stage of learning is then employed to refine the plan. This is guided by the plan, which significantly constrains the parameters governing the robot's actuator movements. This narrows the search space of low-level reinforcement learning that discovers satisficing or optimal values for the parameters.

This learning framework has been applied to locomotion tasks on a multi-tracked robot typical of those designed for urban search and rescue. In addition, it is proposed to use this framework to enable social robots operating in workplace environments to be taught socially desirable skills.

O predavatelju:

Dr Timothy Wiley is currently a Research Associate at the Creative Robotics Lab at Art & Design, and with the AI & Robotics Research Group in the School of Computer Science and Engineering, both at UNSW Sydney.

He has two primary research interests. Firstly, in using autonomous robots as a platform for efficient machine learning, and secondly in investigating explainable artificial intelligence with applications in human-robot interaction. His current research project, in collaboration with Fuji Xerox Japan, intends to improve human-robot interactions on a social robot designed for use in office environments, through online learning techniques that allows a robot to dynamically adapt its behaviour.

Timothy recently received his PhD in Computer Science at UNSW Sydney, focusing in robotics and machine learning. His thesis proposes a hybrid framework for efficiently learning new behaviours for autonomous robots from limited data sets and limited online trial-and-error learning.