A Knowledge Flow Empowered Cognitive Framework for Decision Making With Task-Agnostic Data Regulation

The extreme complexity of many real-world tasks poses considerable challenges to agents' decision making. Most existing models only rely on task-related data for knowledge learning, while ignoring the important influence of potential task-agnostic factors. Effective learning coupled with both task-related and task-agnostic data can strongly enrich the agent's knowledge and improve its decision making performance. Furthermore, many existing learning models simply leverage data to learn knowledge but fail to express the thought process of decision making as humans do, which significantly limits their explanatory capability. To this end, we propose a novel knowledge flow empowered cognitive framework for real-world tasks. To obtain more reliable and trustworthy knowledge, a bottom-up knowledge learning model is developed, which incorporates both task-related data and task-agnostic data for comprehensive knowledge accumulation and value assessment of influencing factors. To demonstrate the thought process of decision making, a top-down knowledge expression model is proposed to coordinate different influencing factors by a knowledge flow structure. Two real-world case studies, including traffic anomaly detection and vehicle following anomaly detection, are introduced, where task-agnostic data are presented for the first time in both tasks. Experimental evaluation demonstrates the strong necessity of incorporating task-agnostic data in knowledge accumulation for real-world tasks and the effectiveness of our cognitive framework.