Multi-Agent Feedback Enabled Neural Networks for Intelligent Communications

Fanglei Sun, Yang Li, Ying Wen, Jingchen Hu, Jun Wang, Yang Yang, Kai Li

Abstract

Modern wireless communication systems increasingly rely on Deep Learning (DL) for complex tasks such as channel modeling and signal detection. However, standard feedforward or basic feedback architectures often lack the "reflective" processing capabilities required to handle severe Inter-Symbol Interference (ISI) and nonlinear distortions. This paper introduces the Multi-Agent Feedback Enabled Neural Network (MAFENN), a framework that models the communication process as a three-player Feedback Stackelberg game. By incorporating a dedicated "Feedbacker" agent to denoise and refine signals, MAFENN achieves significant performance gains—up to 5 dB in nonlinear environments—over traditional and standard DL-based equalizers.

Introduction

While Deep Learning has demonstrated remarkable success in physical layer communications, most existing models operate as simple mapping functions from input to output. These architectures struggle in complex, real-world environments characterized by hardware imperfections (e.g., mixers and amplifiers) and multipath propagation. To address these challenges, we propose a shift from mechanical feedback loops to an intelligent "rethinking" mechanism. By treating feedback as a collaborative task between multiple neural agents, we can emulate the reflective capabilities of biological systems, allowing the network to iteratively refine its internal representations of noisy signals.

The MAFENN Framework

A Three-Player Feedback Stackelberg Game

The core of the MAFENN architecture is a game-theoretic formulation involving three distinct agents. This structure provides a rigorous mathematical foundation for the interaction between encoding, processing, and feedback:

  1. The Encoder (Leader): Responsible for extracting high-level features from both the raw input signal and the refined data provided by the feedback loop.
  2. The Feedbacker (Follower to Encoder, Leader to Processor): A specialized agent that maps latent variables back into the input space, effectively acting as a learned denoiser that reconstructs a "cleaner" version of the signal.
  3. The Processor (Follower): The final stage of the pipeline that performs the specific task, such as symbol identification or signal classification.

By formulating this as a Stackelberg game, we can prove the convergence of the training process toward a stable "Feedback Stackelberg equilibrium," ensuring that the multi-agent interactions remain robust and predictable.

Architectural Implementation (MAFENN-E)

For the specific application of channel equalization, we developed the MAFENN-E model. This architecture utilizes Convolutional Neural Networks (CNNs) to implement matched filtering and handle shift-invariant properties, combined with Long Short-Term Memory (LSTM) units to capture temporal dependencies and model the inverse of the channel's memory.

The training process employs a dual-loss objective function:

  • Cross-Entropy Loss: Optimizes the Processor for classification accuracy.
  • Mean Squared Error (MSE): Ensures the Feedbacker accurately reconstructs the clean, original signal from the noisy input.

Experimental Methodology

Dataset and Channel Models

The framework was evaluated using a synthetic dataset of 11 million randomly generated digital signals (10 million for training, 1 million for validation) using Quadrature Phase Shift Keying (QPSK) modulation. Performance was tested across two primary channel environments:

  1. Linear Multipath: A 10-tap impulse response channel with additive white Gaussian noise (AWGN).
  2. Nonlinear Multipath: A linear channel followed by a nonlinear distortion function to simulate the effects of real-world hardware components.

Feedback Integration

We investigated two methods for integrating feedback: replacement (substituting raw data with feedback) and concatenation (retaining both raw and feedback data). Our results indicate that concatenation is significantly more effective, as it allows the Encoder to compare the noisy physical reality with its refined internal representation.

Results and Discussion

Performance Gains

MAFENN-E demonstrated substantial improvements over both traditional equalizers (e.g., Recursive Least Squares, RLS) and standard DL models (e.g., MLP and CRNN):

  • Linear Scenarios: MAFENN-E achieved a gain of approximately 2 dB at a Symbol Error Rate (SER) of $10^{-3}$ compared to traditional methods.
  • Nonlinear Scenarios: The framework's advantages were most pronounced here, achieving a gain of over 5 dB compared to standard feedback-based DL methods (FB-CRNN) at an SER of $10^{-5}$.

Robustness and Efficiency

In environments where traditional RLS equalizers failed to reach low error floors due to nonlinearities, MAFENN-E maintained high accuracy. Furthermore, we observed that while performance improves with feedback cycles, it typically plateaus after five iterations, suggesting an optimal balance between computational latency and accuracy.

Conclusion and Future Work

MAFENN represents a significant step toward more intelligent, reflective wireless communication systems. By leveraging multi-agent game theory, the framework effectively "rethinks" noisy signals to overcome complex distortions. Future research will focus on reducing the computational complexity of the Stackelberg optimization and extending the MAFENN framework to other domains such as beamforming, resource allocation, and source coding.

Summary updated

Abstract

In the intelligent communication field, deep learning (DL) has attracted much attention due to its strong fitting ability and data-driven learning capability. Compared with the typical DL feedforward network structures, an enhancement structure with direct data feedback have been studied and proved to have better performance than the feedfoward networks. However, due to the above simple feedback methods lack sufficient analysis and learning ability on the feedback data, it is inadequate to deal with more complicated nonlinear systems and therefore the performance is limited for further improvement. In this paper, a novel multi-agent feedback enabled neural network (MAFENN) framework is proposed, consisting of three fully cooperative intelligent agents, which make the framework have stronger feedback learning capabilities and more intelligence on feature abstraction, denoising or generation, etc. Furthermore, the MAFENN frame work is theoretically formulated into a three-player Feedback Stackelberg game, and the game is proved to converge to the Feedback Stackelberg equilibrium. The design of MAFENN framework and algorithm are dedicated to enhance the learning capability of the feedfoward DL networks or their variations with the simple data feedback. To verify the MAFENN framework’s feasibility in wireless communications, a multi-agent MAFENN based equalizer (MAFENN-E) is developed for wireless fading channels with inter-symbol interference (ISI). Experimental results show that when the quadrature phase-shift keying (QPSK) modulation scheme is adopted, the SER performance of our proposed method outperforms that of the traditional equalizers by about 2 dB in linear channels. When in nonlinear channels, the SER performance of our proposed method outperforms that of either traditional or DL based equalizers more significantly, which shows the effectiveness and robustness of our proposal in the complex channel environment.
BibTeX
@article{sun2022multi,
  author = {Fanglei Sun and Yang Li and Ying Wen and Jingchen Hu and Jun Wang and Yang Yang and Kai Li},
  title = {{Multi-Agent Feedback Enabled Neural Networks for Intelligent Communications}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2022},
  doi = {10.1109/twc.2022.3147499},
}
Figure