Skip to main content

AS Seminar: Adaptive Discounted Fully Probabilistic Design of Decision Strategy

Date
Room

Fully probabilistic design of decision strategies (FPD) obtains optimal policy for decision tasks using only probability distributions. Herein lies its main asset. It also covers Markov decision processes (MDP) as a special case. Methods of solving discounted MDP have already been introduced. However, their approximations are computationally more demanding than FPD approximations. Discounted FPD therefore seems more promising. In case of dealing with incomplete knowledge of state transitions, discounted FPD needs to be generalized using Bayesian model estimation. Model variations in time are treated by forgetting. Certainty-equivalence strategy is applied to the final model to decrease dimensionality of the task. Such treatment should make the discounted FPD version more suitable for practical use.

Submitted by neuner on