Troubleshooting with Simultaneous Models

The goal of decision-theoretic troubleshooting is to find a sequence of actions that minimizes the expected cost of repair of a device. If the device is complex then it is convenient to create several Bayesian Networks, each designed to solve a particular problem. At the beginning of a troubleshooting process, it is often necessary to help the user to select the proper model. Complications arise if the user is able to give only a vague description of the problem. In such a case we need to work simultaneously with many troubleshooting models. In this paper we show how models that were originally designed as independent models can be used together while memory space and computational time are kept low. We allow models to be overlapping, i.e., two or more models may contain equivalent troubleshooting steps and/or equivalent problem causes (device faults). We propose a troubleshooting procedure that can be used with many simultaneous models at once. The key that enables us to join the models together is the single fault assumption, which means that there is only one fault causing a device malfunction at a time. 1 SACSO Troubleshooting Approach We start with a review of the SACSO troubleshooting approach proposed for troubleshooting with a single model. The approach was implemented in the HP BATS troubleshooter [2]. The goal of a troubleshooting task is to find and remove the cause of a device malfunction. In case of a complex device, such as for example a laser printer, it is convenient to create several models each designed to solve a particular problem. All original troubleshooting models Mi, i = 1, 2, . . . , N have similar structure. Each model Mi describes relations between a set of repair actions Ai, a set of observations Oi, and a set of causes Ci that can be solved within model Mi. Repair actions are actions that can directly solve the problem, while observations can not solve the problem directly, but may help identify the problem cause. It is assumed that only one cause from Ci can be the cause of a device malfunction at a time. It is often referred to as the single fault assumption. This assumption is reasonable when troubleshooting printing systems and similar man-made devices. Therefore, each cause can be represented as a state ci ∈ Ci of a single cause variable CMi. The state space of each variable CMi is extended by an additional state n.a. This state corresponds to the case when the true cause of the problem is not addressed in model Mi. In other words, CMi = n.a. corresponds to the situation when model Mi does not solve the problem. It is also assumed that actions and observations are independent given the cause. This assumption implies that if the cause of the problem is known then neither the fact that an action failed to solve the problem nor an outcome of a made observation affect the probability of any other action solving the problem. In Fig. 1 an example of two SACSO troubleshooting models is shown.