学术文献库

In decentralized decision-making problems, agents choose their actions based on locally available information and knowledge about decision rules or strategies of other agents. We consider a three-node cascade network with an encoder, a relay and a decoder, having distinct objectives captured by cost functions. In such a cascade network, agents choose their respective strategies sequentially, as a response to the former agent's strategy and in a way to influence the decision of the latter agent in the network. We assume the encoder commits to a strategy before the communication takes place. Upon revelation of the encoding strategy, the relay commits to a strategy and reveals it. The communication starts, the source sequence is drawn and processed by the encoder and relay. Then, the decoder observes a sequences of symbols, updates its Bayesian posterior beliefs accordingly, and takes the optimal action. This is an extension of the Bayesian persuasion problem in the Game Theory literature. In this work, we provide an information-theoretic approach to study the fundamental limit of the strategic communication via three-node cascade network. Our goal is to characterize the optimal strategies of the encoder, the relay and the decoder, and study the asymptotic behavior of the encoder's minimal long-run cost function.