2026-09-20T02:41:42Zhttps://wakespace.lib.wfu.edu/server/oai/request

oai:null:10339/147252026-09-02T11:44:49Zcom_10339_14934col_10339_38132

Saldana, Santiago 2009-08-07T19:11:03Z 2010-06-18T18:57:53Z 2009-08-07T19:11:03Z 2010-06-18T18:57:53Z 2009-08-07T19:11:03Z https://wakespace.lib.wfu.edu/handle/10339/14725 This work examines how individual Partially Observable Markov Decision Process planning performance can be improved through the group-sharing of previous experiences and plans. Individual agents make use of a cache to map between beliefs and appropriate actions. Cache entries are shared between two types of agents using three different protocols: fullBlast, singleQuery, and nonExpert. Expert agents are used to contribute to group knowledge and nonExpert agents make use of group knowledge. The fullBlast and singleQuery protocols follow the teacher-student model, where nonExpert agents make use of the knowledge gained by Expert agents. The fullBlast approach consists of sharing all cache entries at once while the singleQuery variation consists of querying for an entry on an on-demand basis. The nonExpert protocol consists of running agents simultaneously while synchronizing agent knowledge when new states in the world are encountered. Results show that total planning time can be significantly reduced compared to individual agent planning time by synchronizing agent knowledge with group knowledge on an on-demand basis. en_US Wake Forest University Computer Science POMDP Plan Sharing The Effectiveness of Social Plan Sharing in Online Planning in POMDP-type domains Thesis Release the entire work immediately for access worldwide. John, David Turkett, William Jr Pauca, Victor Computer Science open.access