A Hybrid Artificial Intelligence Architecture for Battlefield Information Fusion Paul G. Gonsalves, Gerard J. Rinkus, Subrata K. Das and Nick T. Ton Charles River Analytics, Inc. 725 Concord Avenue, Cambridge, MA 02138 USA Email:{pgonsalves, grinkus, sdas, nton}@cra.com Abstract - The processing of tactical information and the associated situation assessment of the tactical battlefield is a major task for military personnel. Significant effort has been made in countering this challenge with advances in sensor capabilities and enhancements in avionics, electronics and C4I (command, control, communications, computer and intelligence) systems. This rapid evolution must be met with concomitant advances in information fusion and situation assessment. Additionally, a rapid verifiable means is needed in situ for management of sensor and information assets. Here, an on- going effort to develop a hybrid artificial intelligence architecture for battlefield information fusion is described. The architecture incorporates three distinct modules: a low-level information fusion module incorporating a fuzzy expert system manager; a situation assessment module incorporating a fuzzy logic based event detector and a Bayesian belief network component for generating probability measures of situational state; and a fuzzy expert system based module for collection or sensor management. Keywords: Information Fusion, Situation Assessment, Belief Networks, Fuzzy Logic I. Introduction The analysis of intelligence data to generate a comprehensive understanding of all tactical elements within the battlespace and their likely evolution, i.e., to achieve situation awareness is a major task for military personnel. This task naturally overlaps with and benefits from the tasking and management of the sensor/collection assets themselves. Here, we develop a hybrid artificial intelligence (AI) architecture that provides an integrated framework for analysis of information in support of enhanced tactical awareness and needs-based sensor asset management to assist in battlefield intelligence processing. The architecture’s flexibility stems from combining two AI techniques for model-based approximate reasoning: fuzzy logic and the Bayesian belief networks. Information fusion strives to combine information from multiple sources into information that has greater benefit than would have been derived from each of the contributing parts. An obvious analogy exists between fusion and human cognitive processing, in particular, the way humans process multi-sensory information (i.e., sight, sound, smell, etc.) to make inferences regarding the environment. Our hybrid AI battlefield information fusion system uses a coordinated application of two artificial intelligence technologies, fuzzy logic (FL) and Bayesian belief networks (BNs), to the problem of tactical fusion and collection management. Fuzzy logic [1] provides a means of converting low-level imprecise information in non-numerical format into mid-level knowledge units about individual battlespace entities. Belief networks [2] [3] provide a means for constructing and maintaining a hierarchical, probabilistic model linking multiple entities, at various levels, in the context of the overall mission goals, rules of engagement, etc. Evidence gathered incrementally and in real-time first undergoes FL filtering and is then applied to the appropriate node(s) of the BN. This evidence then automatically propagates throughout the BN resulting in revised probability estimates concerning the higher-level tactical situational hypotheses. Experiences from prior research efforts [4][5] have shown that this approach provides an effective solution to the problem and offers a natural framework for encoding complex tactical knowledge. II. System Description Figure 1 illustrates how the overall scope of the hybrid architecture for battlefield information fusion falls within the various levels of fusion [6] and other key components of tactical C4I systems. Information concerning the various entities present in the battlespace, are collected by a variety of sensor or collection assets (JSTARS, AWACS, etc.) and then fused (level one) within the architecture to generate individual target tracks and to classify and characterize targets. The situation assessment (SA) module of the architecture uses this fused track data to generate a probabilistic situational state hypothesis