This paper investigates the problem of coordinating several agents through their actions, focusing on an asymmetric observation structure with two agents. Specifically, one agent knows the past, present, and future realizations of a state that affects a common payoff function, while the other agent either knows the past realizations of nothing about the state. In both cases, the second agent is assumed to have strictly causal observations of the first agent's actions, which enables the two agents to coordinate.These scenarios are applied to distributed power control; the key idea is that a transmitter may embed information about the wireless channel state into its transmit power levels so that an observation of these levels, e.g., the signal-to-interference plus noise ratio, allows the other transmitter to coordinate its power levels. The main contributions of this paper are twofold. First, we provide a characterization of the set of feasible average payoffs when the agents repeatedly take long sequences of actions and the realizations of the system state are i.i.d.. Second, we exploit these results in the context of distributed power control and introduce the concept of coded power control. We carry out an extensive numerical analysis of the benefits of coded power control over alternative power control policies, and highlight a simple yet non-trivial example of a power control code.A central question is to characterize the possible values of the average payoff W N when the agents interact many times, i.e., when N is large. Answering this question in its full generality still appears out of reach, and the present paper settles for a special case with K = 2 agents. Specifically, we assume that Agent 1 has perfect knowledge of the past, current, and future realizations of the state sequence x N 0 , while Agent 2 obtains imperfect and strictly causal observations of Agent 1's actions and possesses either strictly causal or no knowledge of the realizations of the state. Despite these restricting assumptions, one may extract valuable concepts and insights of practical relevance from the present work, which can be extended to the general case of K ≥ 2 agents and arbitrary observation structures.
A. Related workIn most of the literature on agent coordination, including classical team decision problems [4], agents coordinate their actions through dedicated channels, which allow them to signal or communicate with each other without affecting the payoff function. The works most closely related to the present one are [5], [6], in which the authors introduce the notions of empirical and strong coordination to measure agents' ability to coordinate their actions in a network with noiseless dedicated channels. Empirical coordination measures an average coordination over time and requires the joint empirical distribution of the actions to approach a target distribution asymptotically in variational distance; empirical coordination relates to the communication of probability distributions [7] and tools from rate-distortion theor...