Online Privacy Policies: An Assessment of the Fortune Global 100
Jinchang Wang · The Journal of International Information Management · 2005
In many applications of intelligent agents, initially given facts are not sufficient to reach a decision, and more data are needed. In that case, Inference-guiding is needed to identify the missing information and lead inference to a conclusion. This paper presents a new inferenceguiding strategy that selects the key pieces of missing information in such a way that the total cost of acquiring additional information for reaching a conclusion is the lowest. The computational experiments show that the new strategy is more effective and economical than the inference-guiding strategies currently available for the intelligent systems. INTRODUCTION The ‘intelligent agent’, a term in artificial intelligence, refers to a device or a system that can, to some extent, ‘think’ as human beings and ‘act’ rationally. A robot is a typical example of an intelligent agent. An intelligent agent as a node in a computer network handles the information transmitted through the node. Other examples include an automatic real-time control mechanism, a computerized system for disease diagnosing, debugging, or professional training. Comparing to an expert system and a knowledge-based system, which are computer systems that store human’s knowledge and mimic the human’s logic to solve problems in certain domains, an intelligent agent is more self-contained, more autonomous, and more action-oriented. An intelligent agent has three major functions: perceiving the environment, making judgments / decisions, and acting. Percepts of environment are obtained through perceptors such as various sensors, data input channels from a database, and devices for inputs from users. Judgments and decisions are made in the ‘brain’ of the intelligent agent, which contains a knowledge base and an inference engine. The knowledge base stores knowledge. The inference engine is software for doing logical inference. The third function, acting, is to execute the decision or announce the judgment through actuators. An actuator is a device such as an arm of a robot, a control mechanism, and a natural language output device. In many applications of intelligent agents, such as diagnosing, training, and real-time control, initial data about the environment are often incomplete, because of huge amount of the ‘complete’ data and the cost of obtaining them. The cost of obtaining data can be monetary cost or cost of time. To get information from electrocardiograms and CT, for example, would cost money, and to ask a patient questions would cost time. If no solid conclusions can be made due to insufficient data, an intelligent agent should be able to pinpoint the missing data, make hypothesis, collect more information, and prove the hypothesis. In this sense, an intelligent agent should be not only a ‘thinker’ but also an ‘investigator’. Just as a human inspector has to do investigation when initial clues for a case are not sufficient. An intelligent agent for disease diagnosing, which makes judgment on the ailment that a patient may have, must also do investigations by, for example, asking the patient to provide more information on symptoms and selecting some examinations for the patient to do, when the information on hand is not sufficient to get any solid conclusion. The process of ‘investigation’ is called inference-guiding (IG), or question-asking [Wang et al. (1990)], which pursues more information about the environment for further inference. A good IG process would be able to identify a few relevant and key missing data in an efficient way, and lead inference quickly to a conclusion. A bad IG process, on the other hand, would ask irrelevant, costly, and silly questions, and retard the inference process. The function of inference-guiding resides in the inference engine of an intelligent agent. The inference engine should be able to, like human being, figure out what information should be pursued if the current known data are not sufficient. A human inspector ought to be good in not only analysis but also investigation. That is, he J. Wang 2005 Volume 14, Numbers 1 & 2 94 should be good in exploring the meaning of the clues he has by logically linking them, and reaches a conclusion if the clues are sufficient; and he should also be good in investigations, if the clues are not sufficient, by figuring out the key missing data and collecting the data to prove his hypothesis. An inference engine of an intelligent agent thus has two tasks. One is logical inference (or simply inference), which is to make deductions to reveal logical implications of the known information and to make the judgment or decision. Another task is inference-guiding, which is to identify the missing data if given facts are not sufficient to reach any solid conclusion. Figure 1 shows an intelligent agent with inference-guiding function and its interaction with the environment. Figure 1. The intelligent agent with inference guiding function. Logical inference is a subject that has been studied extensively. The Davis-Putnam algorithm [Davis et al. (1960)] and DPLL backtracking algorithm [Davis et al. (1962)] were among the earliest effective algorithms for propositional knowledge bases. Thereafter Robinson developed the full resolution rule [Robinson (1965)]. Due to the close relation between propositional inference and the satisfiability problem (SAT), all the algorithms developed for SAT are actually working for propositional inference either [Gu et al. (1997)]. Modus ponens [Bonissone (1993)] [Awad (1996)] is a deductive rule among implications. Forward chaining (or data-driven) and backward chaining (or goal-driven) [Turban et al. (2001)] are two alternative methods for controlling inference in rule-based intelligent systems, based on which there are many variations and extended technologies, such as the production system [Palopoli et al. (1997)] and deductive database [Ramakrisshnan et al (1995)] [Ullman (1989)]. The Jeroslow-Wang algorithm [Jeroslow et al. (1990)] utilized the integer programming techniques into logical inference. That algorithm was further improved in [Wang (1997)] and [Wang (1998-B)]. Comparing to logical inference, inference-guiding has been less explored. But still some research results have been achieved. EXPERT used pre-listed orderings of rules and questions [Hayes-Roth et al. (1983)]. KAS, a shell over PROSPECT, used both forward and backward chaining, together with a scoring function, for picking more relevant missing data [Duda et al. (1979)]. Mellish’s procedure [Mellish (1985)], using a so-called ‘Alpha-beta pruning technique’, eliminated irrelevant questions for acyclic inference nets. Wang and Vande Vate proved that the problem of identifying the fewest key questions was computationally hard even in a Horn system, and they developed a heuristic algorithm for Horn systems [Wang et al. (1990)]. A cost-effective IG strategy for Horn systems was developed by perceptors knowledge base inference for inference inferenc -guiding actuator Intelligent Agent Environment / user requests for more judgment / decision dat Inference-Guiding for Intelligent Agents Journal of International Technology and Information Management 95 Wang and Triantaphyllou [Wang et al. (1996)]. For the propositional systems, Wang’s IG algorithm in [Wang (1998-A)] aimed at selecting fewer questions. The algorithm in [Wang (2005)] took the cost factor into account. We present a new IG approach in this paper, which is an improvement of the algorithm in [Wang (2005)]. Section II introduces fundamental concepts and terms. Section III discusses the general process of so-called ‘top-levelconclusion oriented inference’ in an intelligent agent. In Section IV, we review three currently existing IG algorithms. The new IG strategy is presented in Section V, and the results of computational experiments shown in Section VI. Section VII discusses the managerial implications and applications of the new IG strategy.