A Maximum Likelihood Method for Estimating Performance in a Rapid Serial Visual Presentation Target-Detection Task

Abstract

In human-agent teams, communications are frequently limited by how quickly the human component can deliver information to the computer-based agents. Treating the human as a sensor can help relax this limitation. As an instance of this, the rapid serial visual presentation target-detection paradigm provides a fast lane for human target-detection information; however, estimating target-detection performance can be challenging when the inter-stimulus interval is short, relative to human response time variability. This difficulty stems from the uncertainty in assigning each response to the correct stimulus image. We developed a maximum likelihood method to estimate the hit rate and false alarm rate that generally outperforms classic heuristic-based approaches and our previously developed regression-based method. Simulations show that this new method provides unbiased and accurate estimates of target-detection performance across a range of true hit rate and false alarm rate values. In light of the improved estimation of hit rates and false alarm rates, this maximum likelihood method would seem the best choice for estimating human target-detection performance.