Cognitive information systems for context-aware decision support

Abstract

Although advancements in technology has allowed a large amount of data to be collected and stored, the task of turning this torrent of raw data into useful information for real time decision making is constantly exceeding our cognitive capacity. While modern Decision Support Systems (DSS) have started to adopt certain aspects of human cognition, such as Situation-Awareness (SA) and Context-Awareness (CA), there is an urgent need for a new breed of advanced information systems that incorporates a road range of cognitive capabilities, including awareness, pro-activeness, reasoning and learning. To address the above challenge, this thesis proposes a framework of cognitive information systems that integrates SA and a multi-agent based inference engine for Context-Aware Decision Support (CaDS). By modeling the situational and contextual factors in the environment explicitly, the system is designed to reduce the cognitive load of the users by providing a combination of functions, including event classification, action recommendation and proactive decision making. To enable learning capability, a self-organizing neural network known as the Fusion Architecture for Learning and Cognition (FALCON) is embedded into the CaDS framework. FALCON has the inherent ability to remain stable as it learns incrementally in real time. This is needed within the CaDS framework to continuously improve the prediction accuracy of the system. Experimental results are reported using a simulated Command and Control (C2) problem domain to illustrate how the CaDS framework is able to reduce the cognitive load of the users and improve the prediction accuracies for option generation. For tapping a variety of knowledge, this thesis presents a systematic procedure for integrating domain knowledge with Reinforcement Learning (RL) using FALCON. To exploit the inserted domain knowledge and the learned knowledge that are inherently distinct, the greedy exploitation and reward vigilance adaptation strategies are proposed to achieve maximal exploitation of the domain knowledge while retaining the flexibility of exploring new knowledge. Our experimental results based on a 1-v-1 PE problem domain have reported improvement to the efficiency of RL using this approach.