AI seminar: Extended Markov tracking with ensemble actions

Speaker: Zinovi Rabinovich, University of Southampton, U.K.

Egocentric Perceptual Control (EPC) is characterised by the optimality criteria and the control task being formulated in terms of the agent's perceptions. The particular perception we concentrate on are the system dynamics. For instance, if we travel along a road, the control task can be formulated as "seeing that the car returns to the road's center" -- notice the perceptual ("seeing that") and the dynamic ("returns to") elements of the task. In Markovian environments an effective solution algorithm to the above control problem has been developed and is called the Extended Markov Tracking (EMT) controller. However, in spite of its relative robustness, the EMT controller is model based and its performance deteriorates if the environment model becomes incoherent. In this work we apply expert ensembles to augment the environment model on-line, and recover up to 95% of the performance lost due to model Incoherence-.