Diagnosis using a first-order stochastic language that learns

We have created a diagnostic/prognostic software tool for the analysis of complex systems, such as monitoring the “running health” of helicopter rotor systems. Although our software is not yet deployed for real-time in-flight diagnosis, we have successfully analyzed the data sets of actual helicopter rotor failures supplied to us by the US Navy. In this paper, we discuss both critical techniques supporting the design of our stochastic diagnostic system as well as issues related to its full deployment. We also present four examples of its use.Our diagnostic system, called DBAYES, is composed of a logic-based, first-order, and Turing-complete set of software tools for stochastic modeling. We use this language for modeling time-series data supplied by sensors on mechanical systems. The inference scheme for these software tools is based on a variant of Pearl’s loopy belief propagation algorithm [Pearl, P. (1988). Probabilistic reasoning in intelligent systems: Networks of plausible inference. San Francisco, CA: Morgan Kaufmann]. Our language contains variables that can capture general classes of situations, events, and relationships. A Turing-complete language is able to reason about potentially infinite classes and situations, similar to the analysis of dynamic Bayesian networks. Since the inference algorithm is based on a variant of loopy belief propagation, the language includes expectation maximization type learning of parameters in the modeled domain. In this paper we briefly present the theoretical foundations for our first-order stochastic language and then demonstrate time-series modeling and learning in the context of fault diagnosis.     

An Evaluation of Explanations of Probabilistic Inference

Providing explanations of the conclusions of decision-support systems can be viewed as presenting inference results in a manner that enhances the user's insight into how these results were obtained. The ability to explain inferences has been demonstrated to be an important factor in making medical decision-support systems acceptable for clinical use. Although many researchers in artificial intelligence have explored the automatic generation of explanations for decision-support systems based on symbolic reasoning, research in automated explanation of probabilistic results has been limited. We present the results of an evaluation study of INSITE, a program that explains the reasoning of decision-support systems based on Bayesian belief networks. In the domain of anesthesia, we compared subjects who had access to a belief network with explanations of the inference results to control subjects who used the same belief network without explanations. We show that, compared to control subjects, the explanation subjects demonstrated greater diagnostic accuracy, were more confident about their conclusions, were more critical of the belief network, and found the presentation of the inference results more clear.     

Parallel belief revision: Revising by sets of formulas

The area of belief revision studies how a rational agent may incorporate new information about a domain into its belief corpus. An agent is characterised by a belief state K, and receives a new item of information α which is to be included among its set of beliefs. Revision then is a function from a belief state and a formula to a new belief state.We propose here a more general framework for belief revision, in which revision is a function from a belief state and a finite set of formulas to a new belief state. In particular, we distinguish revision by the set $\{\alpha ,\beta \}$ from the set $\{\alpha \wedge \beta \}$. This seemingly innocuous change has significant ramifications with respect to iterated belief revision. A problem in approaches to iterated belief revision is that, after first revising by a formula and then by a formula that is inconsistent with the first formula, all information in the original formula is lost.This problem is avoided here in that, in revising by a set of formulas S, the resulting belief state contains not just the information that members of S are believed to be true, but also the counterfactual supposition that if some members of S were later believed to be false, then the remaining members would nonetheless still be believed to be true. Thus if some members of S were in fact later believed to be false, then the other elements of S would still be believed to be true. Hence, we provide a more nuanced approach to belief revision. The general approach, which we call parallel belief revision, is independent of extant approaches to iterated revision. We present first a basic approach to parallel belief revision. Following this we combine the basic approach with an approach due to Jin and Thielscher for iterated revision. Postulates and semantic conditions characterising these approaches are given, and representation results provided. We conclude with a discussion of the possible ramifications of this approach in belief revision in general.     

A concurrent dynamic logic of knowledge,belief and certainty for multi-agent systems

This paper extends the logic of knowledge, belief and certainty from one agent to multi-agent systems, and gives a good combination between logic of knowledge, belief, certainty in multi-agent systems and actions that have concurrent and dynamic properties. Based on it, we present a concurrent dynamic logic of knowledge, belief and certainty for MAS, which is called CDKBC logic. Furthermore, a CDKBC model is given for interpreting this logic. We construct a CDKBC proof system for the logic and show that the proof system is sound and complete, and prove that the validity problem for the system is EXPTIME-complete.     

Dynamic Epistemic Logic for Implicit and Explicit Beliefs

Epistemic logic with its possible worlds semantic model is a powerful framework that allows us to represent an agent’s information not only about propositional facts, but also about her own information. Nevertheless, agents represented in this framework are logically omniscient: their information is closed under logical consequence. This property, useful in some applications, is an unrealistic idealisation in some others. Many proposals to solve this problem focus on weakening the properties of the agent’s information, but some authors have argued that solutions of this kind are not completely adequate because they do not look at the heart of the matter: the actions that allow the agent to reach such omniscient state. Recent works have explored how acts of observation, inference, consideration and forgetting affect an agent’s implicit and explicit knowledge; the present work focuses on acts that affect an agent’s implicit and explicit beliefs. It starts by proposing a framework in which these two notions can be represented, and then it looks into their dynamics, first by reviewing the existing notion of belief revision, and then by introducing a rich framework for representing diverse forms of inference that involve both knowledge and beliefs.     

Partial abductive inference in Bayesian belief networks using a genetic algorithm

Abductive inference in Bayesian belief networks is the process of generating the K most probable configurations given an observed evidence. When we are only interested in a subset of the network's variables, this problem is called partial abductive inference. Both problems are NP-hard, and so exact computation is not always possible. This paper describes an approximate method based on genetic algorithms to perform partial abductive inference. We have tested the algorithm using the alarm network and from the experimental results we can conclude that the algorithm presented here is a good tool to perform this kind of probabilistic reasoning.     

The use of explicit goals for knowledge to guide inference and learning

Combinatorial explosion of inferences has always been a central problem in artificial intelligence. Although the inferences that can be drawn from a reasoner's knowledge and from available inputs is very large (potentially infinite), the inferential resources available to any reasoning system are limited. With limited inferential capacity and very many potential inferences, reasoners must somehow control the process of inference. Not all inferences are equally useful to a given reasoning system. Any reasoning system that has goals (or any form of a utility function) and acts based on its beliefs indirectly assigns utility to its beliefs. Given limits on the process of inference, and variation in the utility of inferences, it is clear that a reasoner ought to draw the inferences that will be most valuable to it. This paper presents an approach to this problem that makes the utility of a (potential) belief an explicit part of the inference process. The method is to generate explicit desires for knowledge. The question of focus of attention is thereby transformed into two related problems: How can explicit desires for knowledge be used to control inference and facilitate resource-constrained goal pursuit in general? and, Where do these desires for knowledge come from? We present a theory of knowledge goals, or desires for knowledge, and their use in the processes of understanding and learning. The theory is illustrated using two case studies, a natural language understanding program that learns by reading novel or unusual newspaper stories, and a differential diagnosis program that improves its accuracy with experience.     

Data Parallel Implementation of Belief Propagation in Factor Graphs on Multi-core Platforms

We investigate data parallel techniques for belief propagation in acyclic factor graphs on multi-core systems. Belief propagation is a key inference algorithm in factor graph, a probabilistic graphical model that has found applications in many domains. In this paper, we explore data parallelism for basic operations over the potential tables in belief propagation. Data parallel techniques for these table operations are developed for shared memory platforms. We then propose a complete belief propagation algorithm using these table operations to perform exact inference in factor graphs. The proposed algorithms are implemented on state-of-the-art multi-socket multi-core systems with additional NUMA-aware optimizations. Our proposed algorithms exhibit good scalability using a representative set of factor graphs. On a four-socket Intel Westmere-EX system with 40 cores, we achieve 39.5 $\times $ speedup for the table operations and 39 $\times $ speedup for the complete algorithm using factor graphs with large potential tables.     

Cooperative Diagnosis and Routing in Fault-Tolerant Multiprocessor Systems

In this note, we consider the problem of fault-tolerant routing in multiprocessor systems when incomplete, or partial, diagnostic information is available. We first define a new type of partial diagnosis, known as k-reachability diagnosis. The overhead for k-reachability diagnosis increases with k, which specifies the radius of diagnostic information maintained by each node. We then present a routing algorithm, known as Algorithm Partial Route, that makes use of k-reachability diagnostic information and allows a trade-off between the amount of diagnostic information and the quality of routing. Partial Route is the first algorithm capable of handling systems of arbitrary topology containing an arbitrary number of faults. The worst-case performance of the algorithm in an n-node system, is shown to be optimal when k = n − 1 and within a factor of 2 of optimal when k = 1. Simulation results on meshes and hypercubes are also presented that show, in the average case, Algorithm Partial Route is nearly optimal for relatively small values of k.     

Signaling game based strategy of intrusion detection in wireless sensor networks

As Wireless Sensor Networks (WSNs) become increasingly popular, it is necessary to require Intrusion Detection System (IDS) available to detect internal malicious sensor nodes. Because sensor nodes have limited capabilities in terms of their computation, communication, and energy, selecting the profitable detection strategy for lowering resources consumption determines whether the IDS can be used practically. In this paper, we adopt the distributed-centralized network in which each sensor node has equipped an IDS agent, but only the IDS agent resided in the Cluster Head (CH) with sufficient energy will launch. Then, we apply the signaling game to construct an Intrusion Detection Game modeling the interactions between a malicious sensor node and a CH-IDS agent, and seek its equilibriums for the optimal detection strategy. We illustrate the stage Intrusion Detection Game at an individual time slot in aspects of its player’s utilities, pure-strategy Bayesian–Nash equilibrium (BNE) and mixed-strategy BNE. Under these BNEs the CH-IDS agent is not always on the Defend strategy, as a result, the power of CH can be saved. As the game evolves, we develop the stage Intrusion Detection Game into a multi-stage dynamic Intrusion Detection Game in which, based on Bayesian rules, the beliefs on the malicious sensor node can be updated. Upon the current belief and the Perfect Bayesian equilibrium (PBE), the best response strategy for the CH-IDS agent can be gained. Afterward, we propose an intrusion detection mechanism and corresponding algorithm. We also study the properties of the multi-stage dynamic Intrusion Detection Game by simulations. The simulation results have shown the effectiveness of the proposed game, thus, the CH-IDS agents are able to select their optimal strategies to defend the malicious sensor nodes’ Attack action.     

Social learning with heterogeneous agents and sequential decision making

In this paper, we consider the problem of social learning in a network of agents where the agents make decisions sequentially by choosing one of two hypotheses on the state of nature. Each agent observes a signal generated according to one of the hypotheses and knows the decisions of all the previous agents in the network. The network contains two types of agents: rational and irrational. A rational agent makes a decision by not only using its private observation but also the decisions of each of the agents which already made decisions. To that end, the agent employs Bayesian theory. An irrational agent makes a decision by ignoring the available information and by randomly choosing the hypothesis. We analyze the asymptotic performance of a system with rational and irrational agents where we study rational agents that use either a deterministic or random decision making policies. We propose a specific random decision making policy that is based on the social belief and the private signals of the agents. We prove that under mild conditions the expected social belief in the true state of nature tends to one if the rational agents use the proposed random policy. In a network with rational agents that use deterministic policy, the conditions for convergence are stricter. We provide simulation results on the studied systems and compare their performance.     

Consistency checking reduced to satisfiability of concepts in terminological systems

We investigate the inference problem in knowledge representation systems of theKl-one family. These systems, also called terminological systems, are equipped with concept languages that are used to express the conceptual knowledge of a problem domain in a structured way. In order to reason with the represented knowledge, terminological systems provide a couple of inference services. In this paper we show that the main reasoning problems in expressive concept languages can be reduced to a particular inference problem, namely checking satisfiability of concepts. This result has two important applications. From a practical point of view, our reduction together with the existence of relatively efficient implementations of satisfiability algorithms strongly simplifies the implementation of inference algorithms in terminological systems. Even from a complexity point of view, the result shows that in the underlying concept language interesting inference problems such as consistency checking or query answering are not harder (in terms of the worst case complexity) than satisfiability checking of concepts.This work has been carried out while the author was an employee of the German Research Center for AI (DFKI GmbH), Saarbrücken, Germany.     

Distributed Bayesian diagnosis for modular assembly systems—A case study

The growing interest in modular and distributed approaches for the design and control of intelligent manufacturing systems gives rise to new challenges. One of the major challenges that have not yet been well addressed is monitoring and diagnosis in distributed manufacturing systems. In this paper we propose the use of a multi-agent Bayesian framework known as Multiply Sectioned Bayesian Networks (MSBNs) as the basis for multi-agent distributed diagnosis in modular assembly systems. We use a close-to-industry case study to demonstrate how MSBNs can be used to build component-based Bayesian sub-models, how to verify the resultant models, and how to compile the multi-agent models into runtime structures to allow consistent multi-agent belief update and inference.     

Manipulating convention emergence using influencer agents

Coordination in open multi-agent systems (MAS) can reduce costs to agents associated with conflicting goals and actions, allowing artificial societies to attain higher levels of aggregate utility. Techniques for increasing coordination typically involve incorporating notions of conventions, namely socially adopted standards of behaviour, at either an agent or system level. As system designers cannot necessarily create high quality conventions a priori, we require an understanding of how agents can dynamically generate, adopt and adapt conventions during their normal interaction processes. Many open MAS domains, such as peer-to-peer and mobile ad-hoc networks, exhibit properties that restrict the application of the mechanisms that are often used, especially those requiring the incorporation of additional components at an agent or society level. In this paper, we use Influencer Agents (IAs) to manipulate convention emergence, which we define as agents with strategies and goals chosen to aid the emergence of high quality conventions in domains characterised by heterogeneous ownership and uniform levels of agent authority. Using the language coordination problem (Steels in Artif Life 2(3):319–392, 1995), we evaluate the effect of IAs on convention emergence in a population. We show that relatively low proportions of IAs can (i) effectively manipulate the emergence of high-quality conventions, and (ii) increase convention adoption and quality. We make no assumptions involving agent mechanism design or internal architecture beyond the usual assumption of rationality. Our results demonstrate the fragility of convention emergence in the presence of malicious or faulty agents that attempt to propagate low quality conventions, and confirm the importance of social network structure in convention adoption.     

The role of assumptions in knowledge engineering

Problem-solving methods are means of describing the inference process of knowledge-based systems. In recent years, a number of these problem-solving methods have been identified that can be reused for building new systems. However, problem-solving methods require specific types of domain knowledge and introduce specific restrictions on the tasks that can be solved by them. These requirements and restrictions are assumptions that play a key role in the reuse of problem-solving methods, in the acquisition of domain knowledge, and in the definition of the problem that can be tackled by knowledge-based systems. In this paper we discuss the different roles assumptions play in the development of knowledge-based systems and provide a survey of assumptions used in diagnostic problem solving. We show how such assumptions introduce targets and bias for goal-driven machine learning and knowledge discovery techniques. © 1998 John Wiley & Sons, Inc.     

Cooperative reinforcement learning in topology-based multi-agent systems

Topology-based multi-agent systems (TMAS), wherein agents interact with one another according to their spatial relationship in a network, are well suited for problems with topological constraints. In a TMAS system, however, each agent may have a different state space, which can be rather large. Consequently, traditional approaches to multi-agent cooperative learning may not be able to scale up with the complexity of the network topology. In this paper, we propose a cooperative learning strategy, under which autonomous agents are assembled in a binary tree formation (BTF). By constraining the interaction between agents, we effectively unify the state space of individual agents and enable policy sharing across agents. Our complexity analysis indicates that multi-agent systems with the BTF have a much smaller state space and a higher level of flexibility, compared with the general form of n-ary (n > 2) tree formation. We have applied the proposed cooperative learning strategy to a class of reinforcement learning agents known as temporal difference-fusion architecture for learning and cognition (TD-FALCON). Comparative experiments based on a generic network routing problem, which is a typical TMAS domain, show that the TD-FALCON BTF teams outperform alternative methods, including TD-FALCON teams in single agent and n-ary tree formation, a Q-learning method based on the table lookup mechanism, as well as a classical linear programming algorithm. Our study further shows that TD-FALCON BTF can adapt and function well under various scales of network complexity and traffic volume in TMAS domains.     

Adaptive diagnosis in distributed systems

Real-time problem diagnosis in large distributed computer systems and networks is a challenging task that requires fast and accurate inferences from potentially huge data volumes. In this paper, we propose a cost-efficient, adaptive diagnostic technique called active probing . Probes are end-to-end test transactions that collect information about the performance of a distributed system. Active probing uses probabilistic reasoning techniques combined with information-theoretic approach, and allows a fast online inference about the current system state via active selection of only a small number of most-informative tests. We demonstrate empirically that the active probing scheme greatly reduces both the number of probes (from 60% to 75% in most of our real-life applications), and the time needed for localizing the problem when compared with nonadaptive (preplanned) probing schemes. We also provide some theoretical results on the complexity of probe selection, and the effect of "noisy" probes on the accuracy of diagnosis. Finally, we discuss how to model the system's dynamics using dynamic Bayesian networks (DBNs), and an efficient approximate approach called sequential multifault; empirical results demonstrate clear advantage of such approaches over "static" techniques that do not handle system's changes.     

Formal approach to fault diagnosis in distributed discrete event systems with OBDD

In this paper, we study the fault diagnosis problem for distributed discrete event systems. The model assumes that the system is composed of distributed components which are modeled in labeled Petri nets and interact with each other via sets of common resources (places). Further, a component’s own access to a common resource is an observable event. Based on the diagnoser approach proposed by Sampath et al., a distributed fault diagnosis algorithm with communication is presented. The distributed algorithm assumes that the local diagnosis process can exchange messages upon the occurrence of observable events. We prove the distributed diagnosis algorithm is correct in the sense that it recovers the same diagnostic information as the centralized diagnosis algorithm. Furthermore, we introduce the ordered binary decision diagrams (OBDD) in order to manage the state explosion problem in state estimation of the system.     

Agent Programming in 3APL

An intriguing and relatively new metaphor in the programming community is that of an intelligent agent. The idea is to view programs as intelligent agents acting on our behalf. By using the metaphor of intelligent agents the programmer views programs as entities which have a mental state consisting of beliefs and goals. The computational behaviour of an agent is explained in terms of the decisions the agent makes on the basis of its mental state. It is assumed that this way of looking at programs may enhance the design and development of complex computational systems.To support this new style of programming, we propose the agent programming language 3APL. 3APL has a clear and formally defined semantics. The operational semantics of the language is defined by means of transition systems. 3APL is a combination of imperative and logic programming. From imperative programming the language inherits the full range of regular programming constructs, including recursive procedures, and a notion of state-based computation. States of agents, however, are belief or knowledge bases, which are different from the usual variable assignments of imperative programming. From logic programming, the language inherits the proof as computation model as a basic means of computation for querying the belief base of an agent. These features are well-understood and provide a solid basis for a structured agent programming language. Moreover, on top of that 3APL agents use so-called practical reasoning rules which extend the familiar recursive rules of imperative programming in several ways. Practical reasoning rules can be used to monitor and revise the goals of an agent, and provide an agent with reflective capabilities.Applying the metaphor of intelligent agents means taking a design stance. From this perspective, a program is taken as an entity with a mental state, which acts pro-actively and reactively, and has reflective capabilities. We illustrate how the metaphor of intelligent agents is supported by the programming language. We also discuss the design of control structures for rule-based agent languages. A control structure provides a solution to the problem of which goals and which rules an agent should select. We provide a concrete and intuitive ordering on the practical reasoning rules on which such a selection mechanism can be based. The ordering is based on the metaphor of intelligent agents. Furthermore, we provide a language with a formal semantics for programming control structures. The main idea is not to integrate this language into the agent language itself, but to provide the facilities for programming control structures at a meta level. The operational semantics is accordingly specified at the meta level, by means of a meta transition system.