An Observation Framework for Multi-Agent Systems

1. An Observation Framework for Multi-Agent Systems Joonas Kesäniemi, Artem Katasonov* and Vagan Terziyan University of Jyväskylä, Finland ICAS, Valencia April 24, 2009 * Artem Katasonov is at present with VTT Technical Research Center of Finland

2. sensor input action output Agent Environment An agent and its social environment • A classic definition of an agent is an encapsulated computer system situated in some environment and capable of flexible, autonomous action in that environment in order to meet its design objectives. • In a multi-agent system, from the point of view of an individual, all the other agents in the system are a part of the environment – social. • The ability to interact with the social environment is as important as the ability to interact with the physical one. It is also the cornerstone of coordination of agents, which is one of the fundamental problems in multi-agent systems.

3. Interacting with a social environment Communication – exchanging messages with other agents. Direct influence – affecting, without communication, properties or available options of other agents, e.g. killing an agent or blocking an agent’s way. Indirect observation of actions – observing the changes in the physical environment that occur due to actions taken by other agents, e.g. a door became open. Direct observation of actions – observing an agent performing an action, e.g. an agent opening a door now. Observation of properties – directly observing a bodily property of an agent, e.g. physical dimensions or what sensors and actuators it has. 3

4. Lack of support for observation In the classic approach to implementing multi-agent systems (AgentSpeak/Jason, 3APL) the environment is explicitly represented (entity separate from agents). Realizing observation is possible. But the approach is not very suitable for distributed systems. In recent years, the focus has shifted towards better support for distributed multi-agent systems - FIPA reference architecture, middleware approach. Middleware supports mainly communication between agents. Environment is implicit. Observation of other agents is practically not supported. A problem: malicious agent joining a multi-agent system. Cannot be registered until noticeable harm is done. Additionally, direct observation can be very useful in many applications. 4

5. Agent Observable Environment (AOE) AOE is a framework that aims at enabling observation, especially of a social environment, in middleware-based multi-agent systems. In the real world, we can observe things due to mediation of rays of light, sound waves, etc. The goal is to create a mechanism somewhat analogical to those mediators – domain-independent and being outside control of the agents (they could intentionally ”show” some things, but some things they could not ”hide”). The framework provides: Possibility to model MAS components with RDF-based observable soft-bodies. Support for both query and publish/subscribe style observation. Ability to restrict the visibility of observable information using observation rules. 5

6. AOE logical architecture

7. Soft-bodies • Every agent has a soft-body - information repository for agent-specific observable information • Although modeled as a part of the agent, it is managed by an environmental service. • The information to the agent’s soft-body can be published both intentionally by the agent itself (e.g. on its state) and automatically by the AOE infrastructure (e.g. on actions of the agent). • Because the repository is managed externally, a suspended or dead agent can still be observable by others. • This repository is semantic – RDF-based. • The global observable environment, which can spawn all the nodes of a distributed agent platform, is then created as the sum of all the soft-bodies in the system. • AOE also allows other entities like environment services to have soft-bodies – to interact with them through the same mechanism. An example is “digital pheromones” service.

8. Observation process and rules • Types of observation in AOE: • Static (query) – observer specifies a pattern and receives a single matching snapshot of observable information. • Dynamic (subscribe) – observer specifies a pattern, and receives a sequence of events. This enables entities to to observe the evolution of something’s state or to receive notifications concerning either addition or deletion of data. • Push - without a prior expressed interest or even consent, an agent is pushed observations. Of course, it is not obliged to react to those or event process them. • Observation rules: • Limit the visibility of the information. • The evaluation context for the rules is always the content of the observable environment. For example, a rule can limit the visibility of the state of an agent X to the agents that can be observed to be part of the same organization as agent X. • Rules can be created by agents and by environment services, rules can have priorities defined.

9. AOE as JADE plug-in • Prototype implementation of AOE framework for the popular JADE middleware platform is done. • Since the role of the AOE is an infrastructural one, it is implemented as a kernel-level service. • Agents publish information on their state deliberately using API provided. • On the other hand, behaviors executed by an agent, as well as messages sent by it, are intercepted using JADE’s event filter mechanism and published to its soft-body automatically.