Catálogo de publicaciones - libros

Effective deployment and utilisation of limited and constrained intelligence resources — including sensors and other sources — is seen as a key issue in modern multinational coalition operations. In this talk, I will examine the application of semantic matchmaking and argumentation technologies to the management of these resources. I will show how ontologies and reasoning can be used to assign sensors and sources to meet the needs of missions, and show how argumentation can support the process of gathering and reasoning about uncertain evidence obtained from sensor probes.

The concept of delegation is central to an understanding of the interactions between agents in cooperative agent problem-solving contexts. In fact, the concept of delegation offers a means for studying the formal connections between mixed-initiative problem-solving, adjustable autonomy and cooperative agent goal achievement. In this paper, we present an exploratory study of the delegation concept grounded in the context of a relatively complex multi-platform Unmanned Aerial Vehicle (UAV) catastrophe assistance scenario, where UAVs must cooperatively scan a geographic region for injured persons. We first present the scenario as a case study, showing how it is instantiated with actual UAV platforms and what a real mission implies in terms of pragmatics. We then take a step back and present a formal theory of delegation based on the use of 2APL and KARO. We then return to the scenario and use the new theory of delegation to formally specify many of the communicative interactions related to delegation used in achieving the goal of cooperative UAV scanning. The development of theory and its empirical evaluation is integrated from the start in order to ensure that the gap between this evolving theory of delegation and its actual use remains closely synchronized as the research progresses. The results presented here may be considered a first iteration of the theory and ideas.

The process of reaching an agreement in a bilateral negotiation to a large extent determines that agreement. The tactics of proposing an offer and the perception of offers made by the other party determine how both parties engage each other and, as a consequence, the kind of agreement they will establish. It thus is important to gain a better understanding of the tactics and potential other factors that play a role in shaping that process. A negotiation, however, is typically judged by the efficiency of the outcome. The process of reaching an outcome has received less attention in literature and the analysis of the negotiation process is typically not as rigorous nor is it based on formal tools. Here we present an outline of a formal toolbox to analyze and study the dynamics of negotiation based on an analysis of the types of moves parties to a negotiation can make while exchanging offers. This toolbox can be used to study both the performance of human negotiators as well as automated negotiation systems.

In this paper we compare state-of-the-art multi-agent reinforcement learning algorithms in a wide variety of games. We consider two types of algorithms: value iteration and policy iteration. Four characteristics are studied: initial conditions, parameter settings, convergence speed, and local versus global convergence. Global convergence is still difficult to achieve in practice, despite existing theoretical guarantees. Multiple visualizations are included to provide a comprehensive insight into the learning dynamics.

Recent advances in peer to peer (P2P) search algorithms have presented viable structured and unstructured approaches for full-text search. We posit that these existing approaches are each best suited for different types of queries. We present PHIRST, the first system to facilitate effective full-text search within P2P networks. PHIRST works by effectively leveraging between the relative strengths of these approaches. Similar to structured approaches, agents first publish terms within their stored documents. However, frequent terms are quickly identified and not exhaustively stored, resulting in a significantly reduction in the system’s storage requirements. During query lookup, agents use unstructured searches to compensate for the lack of fully published terms. Additionally, they explicitly weigh between the costs involved with structured and unstructured approaches, allowing for a significant reduction in query costs. We evaluated the effectiveness of our approach using both real-world and artificial queries. We found that in most situations our approach yields near perfect recall. We discuss the limitations of our system, as well as possible compensatory strategies.

In this paper we propose a multi-agent architecture, made of co-operative information agents, where agents can share with one another their knowledge of the environment and expertise in planning for achieving goals. In particular we consider how through communication such agents can incrementally learn partial and full plans. Such information exchange is particularly useful in the case of situated agents which have diverse abilities and expertise and which have partial views of their environments. It is also useful in the case of agent systems where agents collaborate towards achieving joint or individual goals. We describe an agent model based on abductive logic programming and give detailed protocols and policies of communication. We then define formally what it means for such information exchanges to be , and prove results regarding termination and effectiveness of dialogues based on the formalized policies.

In this paper we study Distributed Data Mining from a Distributed Artificial Intelligence perspective. Very often, databases are very large to be mined. Then Distributed Data Mining can be used for discovering knowledge (rule sets) generated from parts of the entire training data set. This process requires cooperation and coordination between the processors because incon-sistent, incomplete and useless knowledge can be generated, since each processor uses partial data. Cooperation and coordination are important issues in Distributed Artificial Intelligence and can be accomplished with different techniques: planning (centralized, partially distributed and distributed), negotiation, reaction, etc. In this work we discuss a coordination protocol for cooperative learning agents of a MAS developed previously, comparing it conceptually with other learning systems. This cooperative process is hierarchical and works under the coordination of a manager agent. The proposed model aims to select the best rules for integration into the global model without, however, decreasing its accuracy rate. We have also done experiments comparing accuracy and complexity of the knowledge generated by the cooperative agents.

In this paper we investigate methods for analyzing the expected value of adding information in distributed task scheduling problems. As scheduling problems are NP-complete, no polynomial algorithms exist for evaluating the impact a certain constraint, or relaxing the same constraint, will have on the global problem. We present a general approach where local agents can estimate their problem , or how constrained their local subproblem is. This allows these agents to immediately identify many problems which are not constrained, and will not benefit from sending or receiving further information. Next, agents use traditional machine learning methods based on their specific local problem attributes to attempt to identify which of the constrained problems will most benefit from human attention. We evaluated this approach within a distributed cTAEMS scheduling domain and found this approach was overall quite effective.

Traffic management nowadays is one of the key challenges for cities. One drawback of traditional approaches for traffic management is that they do not consider the different valuations of waiting-time reduction of the drivers. These valuations can differ from driver to driver, e.g., drivers who are late for their job interview have a higher valuation of reduced waiting time than individuals driving home from work routinely. This also applies to trucks with urgent load, e.g., as part of a just-in-time production chain. To overcome this problem, we propose a new mechanism for traffic control at intersections called that is valuation-aware. It relies on agent-based driver-assistance systems to allocate the right to cross an intersection. Our evaluation shows that it does yield a significantly higher overall satisfaction.

This paper presents a design of high-performance agent-based intrusion detection system designed for deployment on high-speed network links. To match the speed requirements, wire-speed data acquisition layer is based on hardware-accelerated NetFlow like probe, which provides overview of current network traffic. The data is then processed by detection agents that use heterogenous anomaly detection methods. These methods are correlated by means of trust and reputation models, and the conclusions regarding the maliciousness of individual network flows is presented to the operator via one or more analysis agents, that automatically gather supplementary information about the potentially malicious traffic from remote data sources such as DNS, whois or router configurations. Presented system is designed to help the network operators efficiently identify malicious flows by automating most of the surveillance process.