Abstract:The continuous nonlinear resource allocation problem has broad applications in various fields such as information and communication, transportation, and finance, and this problem often arises as a subproblem in complex programming. Traditional algorithms, however, typically impose monotonicity assumptions, which limits their applicability. In this paper, we propose two novel Lagrangian dual algorithms that solve this problem without such assumptions. By leveraging the problem's convexity, we update the Lagrange multiplier at each iteration based on the current values of the objective and constraint functions. Additionally, we exploit the separability of the problem, allowing the Lagrangian dual problem to be decomposed into n one-dimensional tractable subproblems, significantly enhancing computational efficiency. We also provide a convergence analysis for the proposed algorithms. Extensive numerical experiments demonstrate that the proposed algorithms consistently achieve optimal solutions and significantly enhance computational efficiency, outperforming existing state-of-the-art methods by at least two orders of magnitude in most test problems, particularly in non-quadratic programming problems.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the Continuous Nonlinear Resource Allocation Problem (CONRAP), which is a problem with wide applications in fields such as information communication, transportation, and finance. Traditional algorithms usually impose monotonicity assumptions on the objective function and constraint conditions, which limits their scope of application.
Specifically, this paper mainly solves the following two convex optimization problems:
1. **CONRAP with general inequality constraints**:
\[
\begin{aligned}
& \min_{\mathbf{x}} \phi(\mathbf{x})=\sum_{i = 1}^{n} \phi_i(x_i) \\
& \text{s.t.} \quad g(\mathbf{x})=\sum_{i = 1}^{n} g_i(x_i)\leq b, \\
& \quad x_i\in X_i:=[l_i, u_i], \quad i = 1,\ldots, n,
\end{aligned}
\]
where \(\phi_i:\mathbb{R}\to\mathbb{R}\) and \(g_i:\mathbb{R}\to\mathbb{R}\) are convex and continuously differentiable functions, and \(b\in\mathbb{R}\).
2. **CONRAP with linear equality constraints**:
\[
\begin{aligned}
& \min_{\mathbf{x}} \phi(\mathbf{x})=\sum_{i = 1}^{n} \phi_i(x_i) \\
& \text{s.t.} \quad g(\mathbf{x})=\sum_{i = 1}^{n} a_i x_i = b, \\
& \quad x_i\in [l_i, u_i], \quad i = 1,\ldots, n,
\end{aligned}
\]
where \(a_i\neq 0\) and all have the same sign for \(i = 1,\ldots, n\).
To solve these problems, the paper proposes two new Lagrangian dual algorithms. These algorithms take advantage of the convexity and separability of the problems without imposing monotonicity assumptions on the objective function and constraint conditions. By updating the Lagrangian multipliers and decomposing the Lagrangian dual problem into \(n\) one - dimensional sub - problems, the computational efficiency is significantly improved. In addition, the paper also provides a convergence analysis of the proposed algorithms and verifies their effectiveness and superiority through extensive numerical experiments.