NEURAL NETWORK APPROXIMATION OF MODEL-PREDICTIVE CONTROL FOR A DYNAMIC OBJECT STABILIZATION SYSTEM

Forest fires pose a serious threat to the environment with the potential of causing ecological harm, financial losses, and human casualties. While research suggests that climate change will increase the frequency and severity of these fires, recent developments in deep learning and convolutional neural networks (CNN) have greatly enhanced fire detection techniques and capability. These models can be leveraged by unmanned aerial vehicles (UAVs) to automatically monitor burning areas. However, drones can carry only limited computational and power resources; therefore, on-board computing capabilities are constrained by hardware limitations. This work focuses on the design of segmentation models to identify and localize active burning areas from aerial RGB images processed on limited computing resources. To achieve this goal, the research compares the performance of different variants of the DeepLabv3 neural network model for fire segmentation when trained and tested with the FLAME dataset using a k-fold cross validation approach. Experimental results are compared with U-Net, a benchmark model used with the FLAME dataset, by implementing this model in the same codebase as the DeepLabv3 model. This work demonstrates that a refined version of DeepLabv3, with a MobileNetv2 backbone using pretrained layers and a simplified atrous spatial pyramid pooling (ASPP) module, yields a similar performance to U-Net, with a precision of 87.8% and a recall of 83.2%, while only requiring 20% of the number of parameters involved with the U-Net topology. This significantly reduces memory and power consumption, enabling longer UAV flight duration and reducing the processing overhead associated with sensor input, making it more suitable for deployment on unmanned aerial vehicles. The model’s compact architecture, implemented using TensorFlow and Keras for model design and training, along with OpenCV for image preprocessing, makes it portable and easy to integrate with edge devices such as NVIDIA Jetson boards.

<p>Deep neural networks have achieved tremendous success in many domains (e.g., computer vision~\cite{Alexnet12,vggnet15,fastrcnn15}, speech recognition~\cite{hinton2012deep,dahl2012context}, natural language processing~\cite{dahl2012context,collobert2011natural}, games~\cite{silver2017mastering,silver2016mastering}), however, there are still many challenges in deep learning comunity such as how to speed up training large deep neural networks, how to compress large nerual networks for mobile/embed device without performance loss, how to automatically design the optimal network structures for a certain task, and how to further design the optimal networks with improved performance and certain model size with reduced computation cost.</p> <p>To speed up training large neural networks, we propose to use multinomial sampling for dropout, i.e., sampling features or neurons according to a multinomial distribution with different probabilities for different features/neurons. To exhibit the optimal dropout probabilities, we analyze the shallow learning with multinomial dropout and establish the risk bound for stochastic optimization. By minimizing a sampling dependent factor in the risk bound, we obtain a distribution-dependent dropout with sampling probabilities dependent on the second order statistics of the data distribution. To tackle the issue of evolving distribution of neurons in deep learning, we propose an efficient adaptive dropout (named evolutional dropout) that computes the sampling probabilities on-the-fly from a mini-batch of examples.</p> <p>To compress large neural network structures, we propose a simple yet powerful method for compressing the size of deep Convolutional Neural Networks (CNNs) based on parameter binarization. The striking difference from most previous work on parameter binarization/quantization lies at different treatments of $1\times 1$ convolutions and $k\times k$ convolutions ($k>1$), where we only binarize $k\times k$ convolutions into binary patterns. By doing this, we show that previous deep CNNs such as GoogLeNet and Inception-type Nets can be compressed dramatically with marginal drop in performance. Second, in light of the different functionalities of $1\times 1$ (data projection/transformation) and $k\times k$ convolutions (pattern extraction), we propose a new block structure codenamed the pattern residual block that adds transformed feature maps generated by $1\times 1$ convolutions to the pattern feature maps generated by $k\times k$ convolutions, based on which we design a small network with $\sim 1$ million parameters. Combining with our parameter binarization, we achieve better performance on ImageNet than using similar sized networks including recently released Google MobileNets.</p> <p>To automatically design neural networks, we study how to design a genetic programming approach for optimizing the structure of a CNN for a given task under limited computational resources yet without imposing strong restrictions on the search space. To reduce the computational costs, we propose two general strategies that are observed to be helpful: (i) aggressively selecting strongest individuals for survival and reproduction, and killing weaker individuals at a very early age; (ii) increasing mutation frequency to encourage diversity and faster evolution. The combined strategy with additional optimization techniques allows us to explore a large search space but with affordable computational costs.</p> <p>To further design the optimal networks with improved performance and certain model size under reduced computation cost, we propose an ecologically inspired genetic approach for neural network structure search , that includes two types of succession: primary and secondary succession as well as accelerated extinction. Specifically, we first use primary succession to rapidly evolve a community of poor initialized neural network structures into a more diverse community, followed by a secondary succession stage for fine-grained searching based on the networks from the primary succession. Accelerated extinction is applied in both stages to reduce computational cost. In addition, we also introduce the gene duplication to further utilize the novel block of layers that appeared in the discovered network structure.</p>

In the ocean, submesoscale physical phenomena O(100m) to O(1km) have been reported to play a key role in ocean oxygen ventilation, nutrient supply to the surface ocean, and carbon export, as well as the transfer of energy to larger scales [1]. However, due to limitations in computational resources, current ocean general circulation models are frequently run at resolutions on the order of O(10km) to O(100km) and cannot directly resolve submesoscale turbulence (i.e., subgrid-scale phenomena). Therefore, parameterization schemes are required to simulate these subgrid-scale phenomena.Recent advances in machine learning have triggered the active exploration of data-driven approaches to parameterization for subgrid-scale phenomena that utilize data from observations and simulations. In previous studies [2, 3], the neural network is trained directly using the same variables as the neural network's output, such as viscosity and diffusivity coefficients. However, this approach does not guarantee that the inferred model parameters accurately represent the state of subgrid-scale phenomena they aim to reproduce. We propose a novel parameterization method for estimating diffusivity and viscosity parameters to parameterize subgrid-scale phenomena and have implemented this method in MITgcm, an ocean simulator [4, 5]. This method trains a neural network using the state variables (i.e., velocity fields, potential temperature, and salinity) derived from the simulation results at a resolution that can directly resolve subgrid-scale phenomena. Therefore, unlike previous studies, the diffusivity and viscosity parameters inferred by the trained network can reproduce the global state of subgrid-scale phenomena.The ocean simulator MITgcm is implemented in Fortran, which does not have a built-in package to compute gradients within the neural network, in contrast to deep learning libraries (e.g., PyTorch) like Python. In our previous work [4, 5], we used a quasi-newton optimization method, which does not require computation of these gradients. However, the optimization performance of this method was limited. In this study, we use adjoint code within MITgcm to compute gradients for optimizing neural networks and examine the effect of different optimizers on training performance. AcknowledgmentsThis work used computational resources of supercomputer Fugaku provided by the RIKEN Center for Computational Science through the HPCI System Research Project (Project ID: hp240394).

Many convolutional neural networks(CNNs) have been proposed to solve computer vision tasks such as image classification and image segmentation. However the CNNs usually contain a large number of parameters to determine which consumes very high computation and power resources. Thus, it is difficult to deploy the CNNs on resource-limited devices. Network pruning and network quantization are two main methods to compress the CNNs, researchers often apply these methods individually without considering the relationship between them. In this paper, we explore the coupling relationship between network pruning and quantization, as well as the limits of the current network compression training method. Then we propose a new regularized training method that can combine pruning and quantization within a simple training framework. Experiments show that by using the proposed training framework, the finetune process is not needed anymore and hence we can reduce much time for training a network. The simulation results also show that the performance of the network can over-perform the traditional methods. The proposed framework is suitable for the CNNs deployed in portable devices with limited computational resources and power supply.

Navigation methods for mobile robots need to take various sources of uncertainty into account in order to get robust performance. The ability to improve performance with experience and to adapt to new circumstances is equally important for long-term operation. Real-time constraints, limited computation and memory, as well as the cost of collecting training data also need to be accounted for. In this paper, we discuss our evolving architecture for mobile robot navigation that we use as a test-bed for evaluating methods for dealing with uncertainty in the face of real-time constraints and limited computational resources. The architecture is based on POMDP models that explicitly represent actuator uncertainty, sensor uncertainty, and approximate knowledge of the environment (such as uncertain metric information). Using this model, the robot is able to track its likely location as it navigates through a building. Here, we discuss additions to the architecture: a learning component that allows the robot to improve the POMDP model from experience, and a decision-theoretic path planner that takes into account the expected performance of the robot as well as probabilistic information about the state of the world. A key aspect of both additions is the efficient allocation of computational resources and their practical application to real-world robots.KeywordsPath PlanningRobot NavigationSensor ModelExecution TraceMarkov StateThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1. In this study, neural network (NN) and response surface (RS) models were developed to investigate the response [average daily gain (ADG) and feed efficiency (FE)] of young broiler chickens to dietary protein and lysine. For this purpose, data on their responses to dietary protein and lysine were extracted from the literature and separate NN and RS models were constructed. 2. Comparison between the NN and RS models revealed higher accuracy of prediction with the NN models compared to the RS models. In terms of R 2 values, the NN models developed for both ADG (R 2 = 0.923) and FE (R 2 = 0.904) were far superior to the RS models (R 2 for ADG = 0.511; R 2 for FE = 0.67). This suggests that the NN models can serve as an alternative option to conventional regression approaches including use of RS models. 3. Optimisation of the NN models developed for response to protein and lysine showed that diets containing 220.7 (g/kg of diet) protein and 12.85 (g/kg of diet) lysine maximise ADG, whereas maximum FE is achieved with diets containing 241.3 and 13.12 (g/kg) protein and lysine, respectively. Based on the optimisation results, optimal dietary protein and lysine concentrations for maximum FE in broiler chickens during the starting period are higher than for ADG.

In the evolving realm of autonomous vehicle navigation, the integration of fuzzy logic and neural networks presents a formidable challenge, particularly in the context of real-time, on-the-fly neural network training. This paper addresses the gap in dynamic and adaptable training methods necessary for navigating unpredictable environments with limited computational resources. The primary objective of our study is to empirically validate a hybrid training approach that combines fuzzy logic with back-propagation learning algorithms, aiming to optimize neural network performance under hardware constraints. Our methodology leverages a fuzzy logic trainer to provide initial training sets dynamically, which guide the neural network in adjusting its weights in real time, thus facilitating adaptive learning during navigation tasks. The findings reveal that this integrated approach not only enhances the learning efficiency of neural networks but also significantly improves navigation accuracy in real-time scenarios. These advancements contribute to the field by demonstrating the feasibility of deploying more adaptable and robust autonomous navigation systems, potentially expanding their application in more diverse and challenging environments.

Background: Data mining (DM) is an approach used in extracting valuable information from environmental processes. This research depicts a DM approach used in extracting some information from influent and effluent wastewater characteristic data of a waste stabilization pond (WSP) in Birjand, a city in Eastern Iran. Methods: Multiple regression (MR) and neural network (NN) models were examined using influent characteristics (pH, Biochemical oxygen demand [BOD 5 ], temperature, chemical oxygen demand [COD], total suspended solids [TSS], total dissolved solid [TDS], electrical conductivity [EC] and turbidity) as the regression input vectors. Models were adjusted to input attributes, effluent BOD 5 (BODout) and COD (CODout). The models performances were estimated by 10-fold external cross-validation. An internal 5-fold cross-validation was also used for the training data set in NN model. The models were compared using regression error characteristic (REC) plot and other statistical measures such as relative absolute error (RAE). Sensitivity analysis was also applied to extract useful knowledge from NN model. Results: NN models (with RAE = 78.71 ± 1.16 for BODout and 83.67 ± 1.35 for CODout) and MR models (with RAE = 84.40% ± 1.07 for BODout and 88.07 ± 0.80 for CODout) indicate different performances and the former was better (P < 0.05) for the prediction of both effluent BOD 5 and COD parameters. For the prediction of CODout the NN model with hidden layer size (H) = 4 and decay factor = 0.75 ± 0.03 presented the best predictive results. For BODout the H and decay factor were found to be 4 and 0.73 ± 0.03, respectively. TDS was found as the most descriptive influent wastewater characteristics for the prediction of the WSP performance. The REC plots confirmed the NN model performance superiority for both BOD and COD effluent prediction. Conclusion: Modeling the performance of WSP systems using NN models along with sensitivity analysis can offer better understanding on exploring the most significant parameters for the prediction of system performance. The findings of this study could build the foundation for prospective work on the characterization of WSP operations and optimization of their performances with a view to conducting statistical approaches.

There are challenges to the reliability evaluation for insulated gate bipolar transistors (IGBT) on electric vehicles, such as junction temperature measurement, computational and storage resources. In this paper, a junction temperature estimation approach based on neural network without additional cost is proposed and the lifetime calculation for IGBT using electric vehicle big data is performed. The direct current (DC) voltage, operation current, switching frequency, negative thermal coefficient thermistor (NTC) temperature and IGBT lifetime are inputs. And the junction temperature (T j) is output. With the rain flow counting method, the classified irregular temperatures are brought into the life model for the failure cycles. The fatigue accumulation method is then used to calculate the IGBT lifetime. To solve the limited computational and storage resources of electric vehicle controllers, the operation of IGBT lifetime calculation is running on a big data platform. The lifetime is then transmitted wirelessly to electric vehicles as input for neural network. Thus the junction temperature of IGBT under long-term operating conditions can be accurately estimated. A test platform of the motor controller combined with the vehicle big data server is built for the IGBT accelerated aging test. Subsequently, the IGBT lifetime predictions are derived from the junction temperature estimation by the neural network method and the thermal network method. The experiment shows that the lifetime prediction based on a neural network with big data demonstrates a higher accuracy than that of the thermal network, which improves the reliability evaluation of system.

Federated learning (FL) enables privacy-preserving distributed machine learning by training models on edge client devices using their local data without revealing their raw data. In edge environments, various applications require different neural network models, making it crucial to perform joint training of multiple models on edge devices, known as multi-task FL. While existing multi-task FL approaches enhance resource utilization on edge devices through adaptive resource configuration or client selection, optimizing either of these aspects alone may lead to suboptimality. Therefore, in this paper, we explore a joint client selection and resource configuration method called JCSRC for multi-task FL, aiming to maximize energy efficiency in environments with limited computation and communication resources and heterogeneous client devices. Firstly, we formalize this problem as a mixed-integer nonlinear programming problem considering all these characteristics and prove its NP-hardness. To address this problem, we first design a multi-agent reinforcement learning (MARL)-based client selection method that selects appropriate clients for each task to train their models. The MARL method makes client selection decisions based on the clients’ data quality, energy efficiency, communication, and computation capacity to ensure fast convergence and energy efficiency. Then, we design a particle swarm optimization (PSO)-based resource configuration scheme that configures appropriate computation and bandwidth resources for each task on each client. The PSO scheme makes resource configuration decisions based on theoretically derived optimal CPU frequency and bandwidth to achieve high energy efficiency. Finally, we carry out extensive simulations and testbed-based experiments to validate our proposed JCSRC. The results demonstrate that, in comparison to state-of-the-art solutions, JCSRC can save energy consumption by up to 59% to achieve the target accuracy.

Subject of the research – visual object tracking using various image hashing algorithms for real-time tracking tasks. The goal of this study is to evaluate the tracking success and processing speed of existing and new hashing algorithms for object tracking and to identify the most suitable algorithms to be used under limited computational resources. The objectives of the research include: developing and implementing object tracking based on the aHash, dHash, pHash, mHash, LHash, and LDHash algorithms; comparing the processing speed and accuracy of these methods on the video sequences "OccludedFace2," "David," and "Sylvester"; determining the tracking success rate (TSR) and frames per second (FPS) metrics for each algorithm; analyzing the impact of the search window size, search strategy, and type of hashing on tracking quality, and providing recommendations for their use. The study also explores the trade-off between accuracy and processing speed for each algorithm considering the constraints of limited computational resources. The methods of this study involve testing and evaluating the accuracy and speed of image hashing algorithms on different test video sequences, as well as the use of metrics to determine object similarity using the Hamming distance. The results demonstrate that the aHash and mHash algorithms demonstrate the best accuracy indicators for all hash window sizes, aHash has a higher processing speed, and mHash offers better robustness to changes in lighting and object position. The dHash and pHash algorithms were less effective than the aHash and mHash algorithms due to their sensitivity to changes in scale and rotation. However, perceptual hashing-based methods, such as pHash, are more robust to contrast and blurring. Conclusions. The best hashing algorithms for real-time object-tracking tasks are aHash and mHash. This study underscores the significance of selecting suitable hashing algorithms and search strategies tailored to specific application scenarios and offers possibilities for further optimization.

In order to have a spacecraft orbital motion model, the developer usually tends to use traditional spacecraft orbital motion models. In order to have a good model, the developer resorts to modelling techniques that are more complex. Increasing model complexity increases algorithm execution time and prevents its usage on-board a spacecraft. This is due to real time constraints and limited computational resources usually imposed by the spacecraft on-board computer. This article develops a new method based on neural networks. The developed method allows the developer of spacecraft orbital motion model to alleviate the problem of increasing execution time with increasing model complexity. The positioning error percent of the proposed algorithm should not be more than 5%. The new method effectiveness evaluation is confirmed by the calculations using three test cases. The first test case has been the exact solution of the two-body problem. The second test case has been the Gamma Ray Observatory (GRO) spacecraft verified simulator. The third test case is a spacecraft simulated on the General Mission Analysis Tool (GMAT), which is developed by NASA and private industries. The maximum obtained error percent of the new method in all the test cases is 0.6%, which is considered a very good performance compared to the predefined 5% bound. The average execution time of traditional spacecraft motion models increases approximately 700% as the disturbance model complexity increases. On the contrary, the developed neural networks algorithm has approximately a constant average execution time regardless of the complexity of the disturbance models utilized. This enables using of the newly developed neural networks algorithm with very complex disturbance models without any increase of the computational load.

ARLP: Automatic multi-agent transformer reinforcement learning pruner for one-shot neural network pruning

Identifying urban air pollution sources is essential for public health and environmental sustainability. In this study, we propose a novel hierarchical method for urban air pollution source identification, leveraging deep learning (DL) within an efficient Bayesian inference framework. We rely on observations in the form of two‐dimensional (2D) pollutant concentration distributions, and adopt the Wasserstein distance to model the likelihood probability distribution. The hierarchical nature of the framework stems from the integration of two neural networks (NNs). The first one acts as an emulator that replicates the physical dispersion model to predict future pollution observations recursively over a defined timeframe. These predictions are then used as inputs for the second NN that approximates the distance between predicted and observed pollutant concentration distributions to rapidly compute the likelihood probability. The approach adopts a multi‐model strategy to mitigate the accumulation of errors, particularly those arising from the recursive prediction steps across multiple time intervals, ensuring the reliability of predictions over extended periods. The proposed framework is implemented on graphics processing units (GPUs), enabling scalable computations for real‐world applications and rapid decision making. Through extensive numerical experiments, we demonstrate the suggested method's effectiveness in accurately estimating pollution source parameters, including location, emission rate, and duration, using synthetic observational data. Sensitivity analyses further explore the impact of observational horizons and sampling on solution convergence and accuracy. Numerical results demonstrate robust performances and computational efficiency compared to the conventional approach, particularly in scenarios with limited computational resources and observations availability.

To alleviate communication pressure and terminal resource constraints in mobile edge computing (MEC) networks, this paper proposes a resource allocation optimization method for MEC systems that integrates data compression technology and non-orthogonal multiple access technology. This method considers practical constraints such as terminal device battery capacity and computational resource limitations. By jointly optimizing computational resource allocation, task offloading strategies, and data compression ratios, it constructs an optimization model aimed at minimizing the total task processing latency. Addressing the challenges stemming from the non-convex nature of the problem and the dynamic variations in channel conditions and task requirements, this paper proposes a softmax deep double deterministic policy gradient algorithm, where softmax operator function mitigates both overestimation and underestimation biases inherent in traditional reinforcement learning frameworks, enhancing convergence performance. Utilizing a deep reinforcement learning framework, the algorithm achieves joint decision-making optimization for computational resources, task offloading, and compression ratios, thereby minimizing the total task processing latency while satisfying transmit power and computational resource constraints. Simulation results demonstrate that the proposed scheme exhibits significant advantages over benchmark algorithms in terms of convergence speed and task processing latency.