Event Location: NSH 1305

Abstract: Operator capability is a significant variable in worksite productivity. Evidence indicates that novice operators can be as little as a quarter as productive as expert operators on some tasks. This dramatic variation in skill directly impacts project budget and timeline.

Operator capability generally improves with experience and can be greatly improved through training. However, training incurs direct costs (e.g., fuel, machine wear, and operator wages) and opportunity costs (due to lost work).

Operator training is usually accomplished through oversight by capable humans. There are several drawbacks to this approach. First, human measures of quality are relative: instructors are imperfect and rate capability relative to instructor skill. Hence, while an average operator can train a novice, an expert operator is needed to train an average operator. Second, human measures of quality are of subjective and vary by personnel. A survey of operators will often reveal that there is no single ‘best’ strategy for a particular task. Third, the act of training itself is a skill. An expert machine operator who is also able to effectively share operation knowledge with others is a rare commodity.

The proposed research explores the use of Inverse Optimal Control (IOC) to automatically measure, compare, and improve operator capability. By modeling expert operators as optimal agents solving a Markov Decision Process, IOC can be applied to determine the cost functional used by these experts. To measure novice operator performance, the learned cost functional is leveraged to compare the novice’s reward to expected reward of an expert. On the basis of this comparison, operator actions can be evaluated to detect sub-optimal behavior. This serves as the basis for several approaches to automatically improve operator capability.

Contributions from this thesis will extend the state of the art in several ways. This research constitutes a generational leap in trainer capability. No system exists today that can automatically train site work operators. Novel IOC methods for comparing operators without numerically solving the operator MDP are developed in response to the challenges posed by operator training. Finally, methods for automated operator improvement are examined in the light of active learning.

Committee:William “Red” Whittaker, Chair

George Kantor

William Messner

Miroslaw J. Skibniewski, University of Maryland