Abstract

Distributed multi-agent systems are useful in handling complex, realistic, and larger problems. However, they are susceptible to failures. A multi-agent system has some special characteristics such as sociality, which divides the system into different social groups so that intra-group coupling is much tighter than inter-group coupling. This thesis aims at developing a selective checkpoint based rollback-recovery protocol specially tailored for multi-agent systems, which combines selective checkpointing and logging technologies to achieve the objectives of simple and domino-free rollback and recovery, and fast output commit. A methodology is proposed to provide a step-by-step procedure for designers to follow in deciding where group checkpoints should be inserted. The methodology is demonstrated on an e-trading system implemented on the JADE agent platform. Experiments are conducted to show that the methodology can provide reasonable performance in multi-agent systems.