Emergence of meta-level control in multi-tasking autonomous agents.

One of the basic considerations in the design of autonomous agents is the competence of the agent when presented with many tasks to perform. This thesis demonstrates the need for meta-level control in the task domain of multi-tasking autonomous agents in dynamic environments. It provides the mechanisms to achieve merging of tasks during the execution phase applying methods usually used in opportunistic planning. In a multi-tasking environment, an autonomous agent faces, not only the difficulty of acting and reacting in the external world, but also the burden of performing more than a single task at the same time. This requires that at every point during execution the agent be capable of attending all the tasks it performs. The idea is that an agent has control in its task domain if at any point during execution it can choose its next action out of the set of all tasks it performs. This effects to an emergent meta-level control that is task independent. Two characteristics are important to the success of merging tasks execution: The right decomposition of tasks and the interaction between the tasks. The better the decomposition, the more the agent is able to intermix tasks actions. The more tasks share knowledge, the more the agent benefits from merging; by elimination of redundant actions, by resolution of conflicting actions and by a seemingly parallel execution of multiple tasks. The advantages of the merging mechanism are demonstrated in two task domains. First it is implemented and tested in the blocks world. The results show improved performance in the number of external actions taken by the agent compared to the performance in the same task set without merging. The mechanism is then implemented in a more complex domain which has most of the characteristics of realistic dynamic and unpredictable environments. It is shown that merging of tasks execution in this domain enables the agent to cope with unpredictable events without having to resort to traditional planning techniques.PhDApplied SciencesComputer scienceUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/128928/2/9303725.pd