Sequential Decision Making in Dynamic Systems PDF Download

Author: Yixuan Zhai
Publisher:
ISBN: 9781369343229
Category :
Languages : en
Pages :

Get Book Here

Book Description
We study sequential decision-making problems in the presence of uncertainty in dynamic pricing, intrusion detection, and routing in communication networks. A decision maker is usually able to learn from the feedback (observations) in sequential decision-making problems. We consider designing optimal strategies and analyze their performance. In the first part, we consider a dynamic pricing problem under unknown demand models. We start with a monopoly dynamic pricing problem. In this problem, a seller offers prices to a stream of customers and observes either success or failure in each sale attempt. The underlying demand model is unknown to the seller and can take one of M possible forms. We show that this problem can be formulated as a multi-armed bandit with dependent arms. We propose a dynamic pricing policy based on the likelihood ratio test. It is shown that the proposed policy achieves complete learning, i.e. it offers a bounded regret where regret is defined as the revenue loss with respect to the case with a known demand model. This is in sharp contrast with the logarithmic growing regret in multi-armed bandit with independent arms. Later, we consider an oligopoly dynamic pricing problem with a finite uncertainty of demand models. Besides just considering the learning efficiency, we assume that sellers are individually rational and consider strategies within the set of certain kind of equilibria. We formulate the oligopoly problem as a repeated Bertrand game with incomplete information. Two scenarios are investigated, sellers with equal marginal costs or asymmetric marginal cost. For the scenarios with equal marginal costs, we developed a dynamic pricing strategy called Competitive and Cooperative Demand Learning (CCDL). Under CCDL, all sellers would collude and obtain the same average total profit as a monopoly. The strategy is shown to be a subgame perfect Nash equilibrium and Pareto efficient. We further show that the proposed competitive pricing strategy achieves a bounded regret, where regret is defined as the total expected loss in profit with respect to the ideal scenario of a known demand model. For the scenarios with asymmetric marginal costs, a dynamic pricing strategy called Demand Learning under Collusion (DLC) is developed. If sellers are patient enough, a tactic collusion of a subset of sellers may be formed depending on the marginal costs and underlying demand model. Using the limit of means criterion, DLC is shown to be a subgame-perfect and Pareto-efficient equilibrium. The dynamic pricing strategy offers a bounded regret over an infinite horizon. Using discounting criterion, DLC is shown to be subgame-perfect [epsilon]-equilibrium, [epsilon]-efficient and with an arbitrarily small regret. The dual problem as an infinitely repeated Cournot competition is formulated and the economic efficiency measured by the social welfare is discussed between Bertrand and Cournot formulations. In the second part, we consider an intrusion detection problem and formulate it as a dynamic search of a target located in one of K cells with any fixed number of searches. At each time, one cell is searched, and the search result is subject to false alarms. The objective is a policy that governs the sequential selection of the cells to minimize the error probability of detecting the whereabouts of the target within a fixed time horizon. We show that the optimal search policy is myopic in nature with a simple structure. In the third part, we consider the shortest path routing problem in a communication network with random link costs drawn from unknown distributions. A realization of the total end-to-end cost is obtained when a path is selected for communication. The objective is an online learning algorithm that minimizes the total expected communication cost in the long run. The problem is formulated as a multi-armed bandit problem with dependent arms, and an algorithm based on basis-based learning integrated with a Best Linear Unbiased Estimator (BLUE) is developed.