2016/02/23 by Paul N. Beuchat, Beuchat, Paul N., Angelos Georghiou +3
Computer Science · Engineering · #Adaptive Dynamic Programming Control #FOS: Electrical engineering #Mechanical Circulatory Support Devices #Smart Grid Energy Management #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1602.07273
openalex publication_date 2016/02/23 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We study both the value function and Q-function formulation of the Linear\nProgramming approach to Approximate Dynamic Programming. The approach is\nmodel-based and optimizes over a restricted function space to approximate the\nvalue function or Q-function. Working in the discrete time, continuous space\nsetting, we provide guarantees for the fitting error and online performance of\nthe policy. In particular, the online performance guarantee is obtained by\nanalyzing an iterated version of the greedy policy, and the fitting error\nguarantee by analyzing an iterated version of the Bellman inequality. These\nguarantees complement the existing bounds that appear in the literature. The\nQ-function formulation offers benefits, for example, in decentralized\ncontroller design, however it can lead to computationally demanding\noptimization problems. To alleviate this drawback, we provide a condition that\nsimplifies the formulation, resulting in improved computational times.\n