A Resource Allocation Game and its Equilibrium Strategies

TL;DR

This paper introduces a Bayesian resource allocation game, analyzes its equilibrium strategies for small and large player sets, and proposes algorithms for finding Nash equilibria, including a novel chattering regime.

Contribution

It develops a comprehensive analysis of equilibrium strategies in a resource allocation game, including mean-field and Gaussian approximations, and introduces a construction algorithm for Nash equilibria.

Findings

Nash equilibria include identity and AIF functions with specific switch points.

The mean-field approximation yields an equilibrium with one switch point.

The construction algorithm guarantees finite termination and equilibrium existence.

Abstract

In this paper we propose a Bayesian game to allocate resources. In this game, there are $c$ units of resources to be allocated to $n$ players. Agent $i$ has a demand of $V_i$ units of resources and takes action $X_i$ according to a strategy function $s_i$, \ie $X_i=s_i(V_i)$. Payoffs are setup such that player $i$ is contented with no more than $V_i$ units of resources. We assume that resources are granted to the players on a smallest-request-first and all-or-nothing basis. For this game with two players, we analyze the equilibrium strategy functions mathematically within the family of alternating identity-and-flat (AIF) functions. We show that Nash equilibrium profiles consist of two identity functions, two AIF functions with a common switch point, or two AIF functions with one and three switch points, respectively. For an $n$-player game with a large $n$ and a large $c_n$ of order $O(n)$, we present a mean-field first order approximation and a second-order Gaussian approximation for its equilibrium strategy function. The first-order analysis obtains an equilibrium AIF function with one switch point. In Gaussian analysis of large games, we propose a construction algorithm. This construction algorithm begins in searching within the family of AIF functions. If a gradient conflict condition occurs, the game enters a chattering regime, in which players play a continuous, strictly increasing strategy function that is not an identity nor a flat function. Conceptually one can view the chattering regime as if players alternate between a slope-one strategy and a flat strategy infinitely fast in order to sustain a high payoff. We prove that the construction algorithm always obtains a Nash equilibrium and terminates in a finite number of steps. We present several numerical examples for the two player game as well as the Gaussian model.