Coping with bad agent interaction protocols when monitoring partially observable multiagent systems