The Role of Conversation Policy in Carrying Out Agent Conversations

Abstract

Structured conversation diagrams, or conversation specifications, allow agents to have predictable interactions and achieve predefined information-based goals, but they lack the flexibility needed to function robustly in an unpredictable environment. We propose a mechanism that dynamically combines conversation structures with separately established policies to generate conversations. Policies establish limitations, constraints, and requirements external to specific planned interaction and can be applied to broad sets of activity. Combining a separate policy with a conversation specification simplifies the specification of conversations and allows contextual issues to be dealt with more straightforwardly during agent communication. By following the conversation specification when possible and deferring to the policy in exceptional circumstances, an agent can function predictably under normal situations and still act rationally in abnormal situations. Different conversation policies applied to a given conversation specification can change the nature of the interaction without changing the specification.