Integrating AI and optimization for decision support: a survey

Amitava Dutta
1996 Decision Support Systems  
This project's goal is to create an integrated system methodology for situation awareness via automated decision support for profuse, uncertain, and conflicting data from many sources that overwhelm the human ability to extract the salient features. The following will be developed within the project: i) multiple hybrid inference engines that manage uncertainty, ii) approaches to resolve conflicting results, iii) a decision support system methodology that runs what-ifs, generates geo-spatially
more » ... sociated mental models, and data-mines a data repository to find relationships, iv) an advanced, interactive graphical humancomputer interface for representation of geographical scenarios and hierarchical selection of data sets using rotatable overhead data pyramids, v) the software specification of a data gateway that handles system events and data interchange, and vi) a framework for data definition and data repository organization. Our approach to inferencing under uncertainty will employ, among other, fuzzy belief networks, fuzzy Petri nets, predictive inferencing systems, our new ellipsoidal basis functional link neural networks, and fuzzy classifiers. We will resolve conflicts with weighted voting, quality measures, and a case-base of known situations. The fuzzy belief networks we have developed overcome the NP-hardness of Bayesian belief networks and the states of belief networks we intend to employ will form cases that will be stored in a case-base for growing knowledge. The human-computer interface will extend and apply techniques from geographical information systems, data visualization, and virtual reality to develop highly interactive, "pyramid-themed" presentation devices for regional scenarios capable to show elevations, terrain features, man-made objects, and various sets of data. The interactive display will permit the personnel to use the mouse, the joystick, or a virtual hand for event and data selection. In conflicts where armed force must be applied to protect our national interests, the application of firepower by individual warfighters and artillery along a front line will be replaced in the main by the use of humans in intelligent sensing systems with GPS. These systems call in rapid responses with massive firepower from afar on precisely located targets. This, combined with the availability of sensors on unmanned aerial vehicle and satellite platforms and with the presence of a rich network of communications have caused the fog of war (unclear situations) to become the glare of war: profuse, uncertain, and conflicting data from many sources overwhelm the human ability to extract the usable information for targeting and other decision making purposes. Vast amounts of data must be parsed, interpreted, organized, stored, accessed and patterns must be detected through complex algorithmic processing. Fused sensor and intelligence data must be used in fast inferencing and cross checking methods where the results are presented in a manner that exploits the experience, intuition, and spatial visualization of commanders in interaction with high speed computers. The commanders need specific information at each stage of the overall process of deciding near-optimal allocation of resources over time in an unfolding scenario. Today's high speed computers and specialized algorithms permit an essentially real-time solution, but the many algorithms for specific tasks are disjoint and do not form a cohesive overall solution. Inferencing Methodologies and Problems Inferencing Models. Our investigation is concerned first with inferencing to support a commander's decision making capability in complex situations where uncertainty abounds. Two methods for decision support are operations research (optimization methods) and inferencing (reasoning), for which the advantages and disadvantages of each are well known [1]. The interest in methodologies for inferencing is universally high for decision making in public policy, business, military strategy, health care, industrial process control, and organizational management as well as in complex technical systems. Inferencing comes in two main categories: cognitive and model based. Cognitive methods involve knowledge formats and reasoning paradigms. Traditional methods are rulebased reasoning (RBR) and the more recent case-based reasoning (CBR). Model-based reasoning (MBR) includes Bayesian probability and various linear and nonlinear programming methods, but only the Bayesian methods deal with uncertainty. CBR can provide powerful inferencing if data from events are stored appropriately. Data mining [2] can reveal associations and rules (e.g., which weapons and strategies are most effective against a particular threat). From Heuristic Rules to Bayesian Belief Networks. Rule-based systems (heuristics) employ certainty factors, but they are "fudge factors" rather than being axiomatic. Such systems are increasingly supplanted with Bayesian belief networks (BBNs) that began as influence diagrams in 1921. They were put on a solid theoretical foundation by Pearl's paper [3] in 1988, which led immediately to a flurry of research activity that continues today. Bayesian probabilities have the powerful capability to apply the influences of other variables, e.g., rule P[B|A] can be further influenced by the outcome of another event C that is described by P [B|A,C], as well as backward and forward influences. The Inadequacy of BBNs for Large-Scale Systems. Although BBNs are theoretically attractive and their models have been employed for many purposes [4, 5], their use has been restricted to relatively small tasks. BBNs actually have drastic problems for large systems. First, they are NP-hard for both exact and approximate methods [6]. With K outcomes possible at each of N nodes there are K N combinations of outcomes for the joint probability distribution that must be computed for exact methods. For a moderately small system of 32 bi-valued nodes there are more than 4 billion (2 32 ) probabilities. Dynamic BNNs [7] are more computationally intensive, although we are currently using a sequence of small BNNs over time for inferencing to monitor relatively small situations [8]. Second, the CPTs are difficult to obtain for moderate and larger systems and require data mining of databases to extract the conditional probabilities via frequencies of associations of occurrences. Deleting or adding any node
doi:10.1016/s0167-9236(96)80001-7 fatcat:5wl6ev5vvzbonblrldayscjcte