Application of evolutionary neural model in path optimisation of simulation system

Meng Xu, Cheng Xu, Luyao Pei, Bo Wang, Zhuanzhuan LvInternational Journal of Information and Communication Technology 2025 · DOI

Abstract

This paper introduces a hybrid "evolving neural model" designed to address the complexities of path optimization in dynamic simulation environments, such as digital twins. By integrating an actor-critic reinforcement learning scheme with Evolutionary Field Optimization (EFO), the proposed framework optimizes both network parameters and architecture. Experimental results across diverse datasets—including cooling load prediction and environmental monitoring—demonstrate superior convergence, accuracy, and stability compared to traditional machine learning benchmarks.

Introduction

Path optimization in simulation systems is a critical challenge, particularly within the context of "digital twins." Traditional graphical algorithms, such as grid-based methods, are often too inflexible for dynamic environments, requiring total reconstruction if start or end points change. Conversely, standard machine learning approaches like Deep Q-Networks (DQN) often struggle with convergence speed and stability in highly complex, non-linear scenarios. This research identifies a critical gap: the need to optimize both network parameters (weights) and structure (architecture) simultaneously to ensure models generalize well without overfitting.

The Evolutionary Neural Network (ENN) Framework

To bridge this gap, the researchers propose an Evolutionary Neural Network (ENN) framework. The core innovation lies in its multi-layered optimization strategy that combines the global search capabilities of evolutionary algorithms with the local fine-tuning of gradient-based learning.

Research methodology flowchart for the evolutionary neural model development.

The framework utilizes Evolutionary Field Optimization (EFO) to perform Neural Architecture Search (NAS). Unlike static networks, the ENN treats the network structure as a "growth field," evolving the number of layers and neurons per layer. To ensure valid architectures, the system employs a "neural repair" mechanism that rounds continuous vectors into discrete integers and removes redundant layers.

Actor-Critic Architecture and LSTM Integration

The core of the learning mechanism is an actor-critic reinforcement learning scheme. The "Critic" layer is specifically designed using Long Short-Term Memory (LSTM) units to process time-series data effectively.

Detailed architecture of the LSTM cell used in the Critic layer to manage temporal data.

The inclusion of LSTM units allows the system to "remember" previous states and how they influenced past decisions. This temporal awareness is crucial for navigating complex simulation environments where current decisions depend on past trajectories. By evolving the architecture while the network learns the task—a concept inspired by the Baldwin effect—the model finds nearly optimal complexity faster than a static network would.

Experimental Validation and Results

The study evaluated the model across several distinct domains to prove its versatility, including energy demand prediction (Cooling Load Dataset), systemic risk simulation (REDS Graphs), and environmental monitoring (NOx Estimation Sensor Array). The hybrid model consistently outperformed standard benchmarks:

  • Convergence: In head-to-head tests against Q-learning and DQN, the proposed actor-critic model showed a much steeper decline in loss, reaching stability in significantly fewer epochs (leveling out between 60–100 rounds).
  • Accuracy: In the cooling load assessment, the optimized architecture achieved a Root Mean Square Error (RMSE) of 0.8648, compared to 1.4794 for less optimized versions.
  • Stability: In NOx estimation tasks, the real-valued PWM-EFO variant achieved an $R^2$ of 0.89, outperforming complex-valued variants and showing superior long-term performance.

Conclusion and Future Work

By integrating the gradient-based learning of neural networks with the global search of evolutionary algorithms, this model maintains a superior exploration-exploitation balance. One of the key insights is the use of an objective function to penalize model complexity, pushing the system toward the simplest possible architecture that still solves the problem.

While the model shows great promise, the authors acknowledge the computational intensity of running evolutionary architecture searches alongside backpropagation. Future research will focus on applying this framework to broader Industry 4.0 applications, specifically intelligent transportation systems and large-scale power grid simulations where digital twins must react to real-time sensor data.

Summary updated

Abstract

Inderscience is a global company, a dynamic leading independent journal publisher disseminates the latest research across the broad fields of science, engineering and technology; management, public and business administration; environment, ecological economics and sustainable development; computing, ICT and internet/web services, and related areas.
BibTeX
@article{xu2025application,
  author = {Meng Xu and Cheng Xu and Luyao Pei and Bo Wang and Zhuanzhuan Lv},
  title = {{Application of evolutionary neural model in path optimisation of simulation system}},
  journal = {International Journal of Information and Communication Technology},
  year = {2025},
  doi = {10.1504/ijict.2025.10074323},
}
Figure