Semantic-compensation-based recovery in multi-agent systems

In agent systems, an agent's recovery from, execution problems is often complicated by constraints that are not present in a more traditional distributed, database systems environment. An analysis of agent-related crash recovery issues is presented, and requirements for achieving 'acceptable' agent crash recovery are discussed. Motivated by this analysis, a novel approach to managing agent recovery is presented. It utilises an event-and task-driven model for employing semantic compensation; task retries, and checkpointing. The compensation/retry model requires a situated model of action and failure, and provides the agent with an emergent unified, treatment of both crash recovery and run-time failure-handling. This approach helps the agent to recover acceptably from crashes and execution problems; improve system predictability; manage inter-task dependencies; and address the way in which exogenous events or crashes can trigger the need for a re-decomposition of a task. Agent architecture is then presented, which uses pair processing to leverage these recovery techniques and increase the agent's availability on crash restart.