From Sigmoid Power Control Algorithm to Hopfield-like Neural Networks: "SIR"-Balancing Sigmoid-Based Networks- Part II: Discrete Time

TL;DR

This paper introduces a discrete-time Sigmoid-based SIR balancing neural network, demonstrating finite-step convergence and drawing parallels to Hopfield networks and power control algorithms, with a new 1-bit variant for simplified implementation.

Contribution

It develops a discrete-time SIR-balancing neural network with proven finite-step convergence and establishes analogies to Hopfield networks and power control algorithms, including a 1-bit version.

Findings

The D-Sgm"SIR"NN converges to zero error in finite steps.

The network's equilibrium points correspond to prototype vectors.

A 1-bit Sign"SIR"NN network is also proposed.

Abstract

In the first part in [12], we present and analyse a Sigmoid-based "Signal-to-Interference Ratio, (SIR)" balancing dynamic network, called Sgm"SIR"NN, which exhibits similar properties as traditional Hopfield NN does, in continuous time. In this second part, we present the corresponding network in discrete time: We show that in the proposed discrete-time network, called D-Sgm"SIR"NN, the defined error vector approaches to zero in a finite step in both synchronous and asynchronous work modes. Our investigations show that i) Establishing an analogy to the distributed (sigmoid) power control algorithm in [10] and [11] if the defined fictitious "SIR" is equal to 1 at the converged eqiulibrium point, then it is one of the prototype vectors. ii) The D-Sgm"SIR"NN exhibits similar features as discrete-time Hopfield NN does. iii) Establishing an analogy to the traditional 1-bit fixed-step power control algorithm, the corresponding "1-bit" network, called Sign"SIR"NN network, is also presented.