Human-Machine Task Allocation Based on Predicate Confidence and Dynamic Cost

Machine learning (ML) models are increasingly being used in application domains that often involve working together with human experts. Effective task allocation between humans and machines is crucial for optimizing performance and maximizing overall efficiency. We propose a novel human-machine task allocation method based on predictive confidence and dynamic cost, aiming to reduce human and machine costs without compromising task accuracy. We first define the task allocation problem, which considers the confidence of humans and machines, the cost of human and machine, as well as task rewards. Our method breaks through the limitations of traditional approaches that assign tasks to single executors only. Instead, it dynamically allocates tasks to human, machine, or both in collaboration to maximize the total expected reward. We have developed an integrated framework that includes key modules such as cost calculation, task allocation, and result fusion. This framework allows us to flexibly determine the task executors based on the specific conditions of each task. Experimental results demonstrate that the proposed method significantly improves the total expected reward compared to multiple baseline schemes, while ensuring task accuracy. Specifically, the proposed method achieves an 8 -fold increase in total expected reward on the CIFAR-10H dataset and approximately a 3-fold increase on the ImageNet-16H dataset.