CISA Seminar: Michael Anslow In many real-world multiagent systems domains, agents start out with some initial conceptualisation of the domain, but discover new entities as they interact with the environment. Given that such new entities are assigned labels at runtime (in an artificial language), the opportunities for prior agreement on ontologies among cooperative agents are limited, and agents with heterogeneous local ontologies have to learn to map terms other agents use to their local ontologies at runtime. We focus on the problem of 'alignment adoption' in this setting. Agents have available to them correspondences between ontologies that may be incorrect and/or incomplete and a reward signal that is indicative of task performance. Using this information, agents must explore/exploit possible alignments to improve task performance. We pose alignment adoption as a multi-armed bandit problem, where known correspondences are used to bias exploration/exploitation of alignments. *This is work in progress towards a conference paper submission* Jan 18 2016 14.00 - 15.00 CISA Seminar: Michael Anslow Finding high-utility alignments in a task context IF 2.33
CISA Seminar: Michael Anslow In many real-world multiagent systems domains, agents start out with some initial conceptualisation of the domain, but discover new entities as they interact with the environment. Given that such new entities are assigned labels at runtime (in an artificial language), the opportunities for prior agreement on ontologies among cooperative agents are limited, and agents with heterogeneous local ontologies have to learn to map terms other agents use to their local ontologies at runtime. We focus on the problem of 'alignment adoption' in this setting. Agents have available to them correspondences between ontologies that may be incorrect and/or incomplete and a reward signal that is indicative of task performance. Using this information, agents must explore/exploit possible alignments to improve task performance. We pose alignment adoption as a multi-armed bandit problem, where known correspondences are used to bias exploration/exploitation of alignments. *This is work in progress towards a conference paper submission* Jan 18 2016 14.00 - 15.00 CISA Seminar: Michael Anslow Finding high-utility alignments in a task context IF 2.33
Jan 18 2016 14.00 - 15.00 CISA Seminar: Michael Anslow Finding high-utility alignments in a task context
