Belief revision, non-monotonic reasoning and secrecy for epistemic agents

Abstract

Software agents are increasingly used to handle the information of individuals and companies. They also exchange this information with other software agents and humans. This raises the need for sophisticated methods of such agents to represent information, to change it, reason with it, and to protect it. We consider these needs for communicating autonomous agents with incomplete information in a partially observable, dynamic environment. The protection of secret information requires the agents to consider the information of agents, and the possible inferences of these. Further, they have to keep track of this information, and they have to anticipate the effects of their actions. In our considered setting the preservation of secrecy is not always possible. Consequently, an agent has to be able to evaluate and minimize the degree of violation of secrecy. Incomplete information calls for non-monotonic logics, which allow to draw tentative conclusions. A dynamic environment calls for operators that change the information of the agent when new information is received. We develop a general framework of agents that represent their information by logical knowledge representation formalisms with the aim to integrate and combine methods for non-monotonic reasoning, for belief change, and methods to protect secret information. For the integration of belief change theory, we develop new change operators that make use of non-monotonic logic in the change process, and new operators for non-monotonic formalisms. We formally prove their adherence to the quality standards taken and adapted from belief revision theory. Based on the resulting framework we develop a formal framework for secrecy aware agents that meet the requirements described above. We consider different settings for secrecy and analyze requirements to preserve secrecy. For the protection of secrecy we elaborate on change operations and the evaluation of actions with respect to secrecy, both declaratively and by providing constructive approaches. We formally prove the adherence of the constructions to the declarative specifications. Further, we develop concrete agent instances of our framework building on and extending the well known BDI agent model. We build complete secrecy aware agents that use our extended BDI model and answer set programming for knowledge representation and reasoning. For the implementation of our agents we developed Angerona, a Java multiagent development framework. It provides a general framework for developing epistemic agents and implements most of the approaches presented in this thesis.