Jan Tožička

Adversarial behavior in multi-agent systems

Adversariality of the agents with respect to the multi-agent system can be a serious issue in the design of open multi-agent systems. Until now, many incoherent definitions of such behavior were used, preventing the consolidation of the knowledge about the domain. By basing ourselves on the valid and accepted results from economics, law and conflict theory, we propose a consistent definition of adversariality in the multi-agent systems and discuss the characteristics of the behavior that falls into this definition.

Market-Inspired approach to collaborative learning

The paper describes a decentralized peer-to-peer multi-agent learning method based on inductive logic programming and knowledge trading. The method uses first-order logic for model representation. This enables flexible sharing of learned knowledge at different levels of abstraction as well as seamless integration of models created by other agents. A market-inspired mechanism involving knowledge trading is used for inter-agent coordination. This allows for decentralized coordination of learning activity without the need for a central control element. In addition, agents can participate in collaborative learning while pursuing their individual goals and maintaining full control over the disclosure of their private information. Several different types of agents differing in the level and form of knowledge exchange are considered. The mechanism is evaluated using a set of performance criteria on several scenarios in a realistic logistic domain extended with adversary behavior. The results show that using the proposed method agents can collaboratively learn properties of their environment, and consequently significantly improve their operation.

Solving inaccessibility in multi-agent systems by mobile middle-agents

This paper analyzes the problem of communication inaccessibility in the multi-agent systems. Apart from listing the main reasons for existence of communication inaccessibility, it suggest a complete inaccessibility model and metrics. Various inaccessibility solutions are presented, together with their applicability in the environments with different degrees of inaccessibility, as verified by experiments. Major contribution of this paper is in suggesting a novel solution to solving inaccessibility based on community of autonomous, migrating middle-agents. A distributed algorithm for dynamic allocation of the middle-agents in a network of partially inaccessible agents is proposed. The suggested solution is supported by a set of experiments comparing the distributed and centralized algorithm for middle-agents allocation.

Agent Methods for Network Intrusion Detection and Response

While the need to build the Intrusion Detection Systems (IDS) based on on a distributed and cooperative (P2P) paradigm is being generally acknowledged, the field has been disconnected from the recent advances in the multi-agent research, most notably the field of trust modeling. Our contribution reviews recent implementations of IDS systems and presents them from an agent research perspective. We also identify the opportunities where the agent approaches can be successfully used. Agent techniques can make the IDS more adaptive, scalable and reliable while increasing their autonomy and reducing the maintenance requirements. Besides trust modeling, we propose that the distributed decision-making and planning techniques can be used to shorten the detection-response loop, making the system more robust while facing worm attacks.

Quantifying Privacy Leakage in Multi-Agent Planning

Multi-agent planning using MA-STRIPS–related models is often motivated by the preservation of private information. Such a motivation is not only natural for multi-agent systems but also is one of the main reasons multi-agent planning problems cannot be solved with a centralized approach. Although the motivation is common in the literature, the formal treatment of privacy is often missing. In this article, we expand on a privacy measure based on information leakage introduced in previous work, where the leaked information is measured in terms of transition systems represented by the public part of the problem with regard to the information obtained during the planning process. Moreover, we present a general approach to computing privacy leakage of search-based multi-agent planners by utilizing search-tree reconstruction and classification of leaked superfluous information about the applicability of actions. Finally, we present an analysis of the privacy leakage of two well-known algorithms—multi-agent forward search (MAFS) and Secure-MAFS—both in general and on a particular example. The results of the analysis show that Secure-MAFS leaks less information than MAFS.

Secure Multi-Agent Planning

Multi-agent planning using MA-STRIPS-related models is often motivated by the preservation of private information. Such motivation is not only natural for multi-agent systems, but is one of the main reasons, why multi-agent planning problems cannot be solved centrally. Although the motivation is common in the literature, formal treatment of privacy is mostly missing. An exception is a definition of two extreme concepts, weak and strong privacy. In this paper, we first analyze privacy leakage in the terms of secure Multi-Party Computation and Quantitative Information Flow. Then, we follow by analyzing privacy leakage of the most common MAP paradigms. Finally, we propose a new theoretical class of secure MAP algorithms and show how the existing techniques can be modified in order to fall in the proposed class.

Towards formal model of adversarial action in multi-agent systems

Detecting and preventing the adversarial action of an agent with respect to the community of agents can be a serious issue in the design of open multi-agent systems. This task is severely domain dependent and it is hard to find a general solution. This contribution presents a utility based model of adversarial action and analyses few properties of adversarial behavior in multi-agent systems. Potential use and important drawbacks of this model are discussed in the paper.

Meta-reasoning methods for agent's intention modelling

Intention modelling in self-interested and adversarial communities of agents is a challenging issue. This contribution discusses the role of modelling and meta-reasoning in intention modelling. The formal model of deductive and inductive meta-reasoning is presented and supported by experimental implementations. This research has been motivated by the problem of intention detection in semi-collaborative multi-agent system for OOTW (Operation Other Than War).

Airports for agents: an open MAS infrastructure for mobile agents

Airports for Agents (AA) is an implemented distributed multi-agent infrastructure designed for dynamic and unstable Internet environment. The infrastructure consists of platforms called Airports that enable agents to communicate together and to be persistent. Furthermore, the Airports allow agents to migrate trough the system and to use local resources. Any Airport can host any agent from the network, therefore we considered high requirements for the security. Network of Airports can dynamically change in the time as new Airports connect to the system, or disconnect. We designed distributed stochastic algorithm keeping the system connected, because AA has no central element. The agent migration brings a communication problem known in the field of distributed systems: where to find the agent I have been communicated with, previously, while he changed his location (platfrom)? We present Kept Connection as a transparent solution of this problem. System is designed to be distributed over large amount of computers.

Recursive Reductions of Action Dependencies for Coordination-Based Multiagent Planning

Currently the most efficient distributed multiagent planning scheme for deterministic models is based on coordination of local agents’ plans. In such a scheme, behavior of other agents is modeled using projections of their actions stripped of all private information. The planning scheme does not require any additional information, however using such can be beneficial for planning efficiency. Dependencies among the projected public actions caused by sequences of local private actions represent one particular type of such information.