Learning to Estimate: A Case-Based Approach to Task Execution Prediction

A system that controls a team of autonomous vehicles should be able to accurately predict the expected outcomes of various subtasks. For example, this may involve estimating how well a vehicle will perform when searching a designated area. We present CBE, a case-based estimation algorithm, and apply it to the task of predicting the performance of autonomous vehicles using simulators of varying fidelity and past performance. Since there are costs to evaluating the performance in simulators (i.e., higher fidelity simulators are more computationally expensive) and in deployment (i.e., potential human injury and deployment expenses), CBE uses a variant of local linear regression to estimate values that cannot be directly evaluated, and incrementally revises its case base. We empirically evaluate CBE on Humanitarian Assistance / Disaster Relief (HA/DR) scenarios and show it to be more accurate than several baselines and more efficient than using a low fidelity simulator.