Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

LSTM-SARIMA Based Prediction Method for Environmental Quality in Enclosed Poultry House

Sun, 03 May 2026 00:00:00 +0000

Closed-type poultry houses facilitate consistent output by ensuring a steady microenvironment conducive to optimal avian growth. Nevertheless, numerous farms continue to depend on manual oversight of temperature, humidity, and ammonia levels, resulting in delayed reactions, diminished productivity, and heightened environmental stress on poultry. These constraints underscore the necessity for predictive and automated systems that can monitor and forecast environmental variables in real time. Prior research indicates that LSTM networks are proficient in nonlinear time-series forecasting nonetheless, when used in isolation, LSTM models encounter difficulties in capturing linear seasonal patterns and long-term trends present in chicken house environmental data. This research presents a hybrid forecasting framework that combines LSTM and SARIMA models to concurrently represent nonlinear temporal dependencies and linear seasonal components. Environmental metrics such as temperature, soil moisture, and ammonia concentration were acquired using SHT31, Soil Moisture, and MQ137 sensors, processed using a Python-Flask backend, saved in MongoDB, and visualized through a cross-platform Flutter-based web interface. Experimental findings indicate that the proposed LSTM–SARIMA model exhibits robust predictive efficacy, with MAE = 0.62, MSE = 0.55, RMSE = 0.58, MAPE = 7.89%, and R² = 0.86. The findings demonstrate that the suggested method efficiently facilitates early warning systems and real-time microclimate evaluation, allowing for expedited environmental management measures and minimizing production losses due to unstable poultry house conditions.

Identification of BSR Disease in Oil Palm from UAV Imagery Using CNN and SCNN Approaches

Zakia Azzahro, Rahmadwati, Angger Abdul Razak, Amrul Faruq — Sun, 03 May 2026 00:00:00 +0000

Basal Stem Rot (BSR) disease caused by Ganoderma boninense is a major threat to oil palm productivity due to its destructive nature and the challenges associated with early-stage detection. To support sustainable production and mitigate significant yield losses, a system capable of identifying tree conditions into healthy and infected categories is required. In this study, two deep learning approaches, CNN and SCNN, are applied to identify oil palm conditions based on UAV-derived imagery. While CNN is widely used for image-based detection tasks due to its ability to extract relevant visual representations, it is prone to overfitting during training, therefore SCNN is employed to address this issue by leveraging image similarity comparison. Experimental results show that both methods achieve high accuracy, with SCNN outperforming CNN by achieving an accuracy of 96.48%, compared to 95.644%. The superior performance of SCNN demonstrates its sensitivity to subtle visual differences between healthy and early-stage infected trees, enabling more reliable models. Thus, SCNN is considered more optimal for detection oil palm conditions and contributes to reducing overfitting, resulting in improved model stability.

An Adaptive Cross-Tied Interconnection for Shaded PV Arrays: A Mathematical Analysis for Efficiency Enhancement

Sun, 26 Apr 2026 00:00:00 +0000

This study investigates the Adaptive Cross-Tied Interconnection (ACTI) configuration to improve the power output efficiency of photovoltaic (PV) arrays operating under partial shade conditions. The objective of this study is to develop a mathematical formulation that describes the behavior of ACTI compared to the conventional Series-Parallel (SP) configuration. Mathematical modeling is used to analyze the current distribution, voltage relationship, and the effect of shading patterns on the total output power. Simulations are performed using MATLAB/Simulink to verify the theoretical analysis results. This adaptive configuration dynamically adjusts the cross-tied based on the illumination intensity data, thus balancing the current between the shaded and normal modules. The results show that ACTI successfully reduces current mismatch losses and increases the output power without increasing circuit complexity. In a 3x3 PV array, the ACTI configuration yields a power increase of up to 48% compared to the SP configuration. In a 5x5 array, the efficiency increase ranges from 2% to 6%, depending on the shading pattern. The adaptive switching strategy maintains the current flow stability and produces a smoother power-voltage curve, allowing faster and more accurate tracking of the global maximum power point. These results demonstrate that ACTI provides an efficient, economical, and mathematically sound solution for improving the performance of PV systems under non-uniform irradiation conditions

Comparison of Nutrient and pH Control in NFT Hydroponic Plants for Coupled and Decoupled Methods

Ina Rahmawati Putri, Bambang Siswojo, Mochammad Rusli — Sun, 26 Apr 2026 00:00:00 +0000

PH and TDS were critical parameters in hydroponic systems that directly influenced nutrient absorption and plant growth. This study developed an automatic nutrient solution control system for NFT hydroponics using a Proportional-Integral-Derivative (PID) controller with coupled and decoupled approaches. The system employed a DFRobot Gravity: Analog TDS sensor to measure TDS, an Electrode Probe pH-4502C to monitor pH, and an Arduino Uno microcontroller to regulate peristaltic pumps in real time. Lettuce was used as the test crop, requiring 550 ppm TDS and pH 6.5. System performance was evaluated through MATLAB Simulink simulations and hardware implementation based on rise time, settling time, overshoot, and steady-state error. The simulation results showed that the coupled method had slightly faster rise time and settling time compared to the decoupled method, whereas the decoupled method had less overshoot than the coupled. The hardware test showed that the decoupled method performed better, with a pH rise time of 8.34 s, a settling time of 11 s, an overshoot of 10%, and a steady-state error of 0.90%, as well as a TDS rise time of 30.7 s, a settling time of 36 s, an overshoot of 4.36%, and a steady-state error of 0.60%. In contrast, the coupled method exhibited slower responses, longer settling times, and higher steady-state errors. Overall, the decoupled method proved more effective and reliable in maintaining pH and TDS stability, showing strong potential to enhance the efficiency and robustness of NFT hydroponic control systems.

An Adaptive Swarm Clustering Algorithm for Game AI Based on Reinforcement Learning Godot and Particle Swarm Optimization (RLGPSO)

Sun, 26 Apr 2026 00:00:00 +0000

Managing extensive agent swarms presents a significant difficulty in dynamic, real-time situations, especially in gaming artificial intelligence, such as real-time strategy. Traditional Particle Swarm Optimization (PSO) techniques, while effective for optimization tasks, often exhibit suboptimal convergence and inadequate flexibility in complex, demanding situations. This study introduces an innovative hybrid approach that integrates Reinforcement Learning (RL) with PSO to create an adaptive swarm clustering system. This approach employs a Deep Deterministic Policy Gradient (DDPG) agent to dynamically modify PSO parameters, enabling the swarm to adeptly maneuver and cluster within a procedurally generated 2D simulation environment featuring physical obstacles, in contrast to earlier studies that depend on static mathematical benchmarks. A rigorous quantitative analysis using Mixed Linear Model Regression (MLMR) demonstrates that this hybrid method significantly and statistically outperforms conventional, manually tuned PSO in terms of convergence time and diversity value. For example, the RLGPSO model achieved an 11.46% reduction in convergence time on high-complexity maps, a result confirmed as statistically significant with a p-value of 0.002 from the MLMR analysis. This study offers a pragmatic approach for the implementation of intelligent, self-organizing agent swarms, directly applicable to improving the realism and efficacy of present-day gaming AI.

Maleo Emotion Audio Dataset Indonesia For Emotion Classification

Sun, 26 Apr 2026 00:00:00 +0000

The limited availability of voice emotion datasets in Indonesian poses a challenge in the development of Speech Emotion Recognition (SER) systems, even though the need for such systems continues to grow in various sectors such as customer service, education, and human-computer interaction. To address this challenge, this study developed the Maleo Emotion Audio Dataset, a collection of three-second audio clips labeled with seven emotion categories: angry, neutral, disgusted, sad, happy, afraid, and surprised. The data was collected from the YouTube platform, and the Maleo Emotion Dataset is available at https://huggingface.co/datasets/maleo-ai/maleo-emotion. It was processed through preprocessing, feature extraction, and augmentation stages. The five main features extracted include Zero Crossing Rate, energy, Mel-Frequency Cepstral Coefficients (MFCC), spectral roll-off, and spectral flux. To enhance generalization, augmentation techniques such as pitch shifting, noise injection, and time stretching were applied. The classification model was built using a Convolutional Neural Network (CNN) architecture with TensorFlow-based implementation. Evaluation showed that the model achieved 94.48% accuracy on the test data, with balanced performance across all emotion categories. These results demonstrate that the developed dataset and model architecture have high capability in effectively recognizing emotions from Indonesian speech in a locally relevant context.

Evaluating Synonym Augmentation Impact on SBERT Performance for Indonesian Social Media Style Classification

Sun, 03 May 2026 00:00:00 +0000

Language on social media reflected the identity and characteristics of its users, including differences in language style between generations. Millennials and Generation Z were two dominant demographic groups in digital communication that exhibited linguistic variations, which often caused gaps in understanding during online interactions. Variations in language structure and expression posed challenges in understanding the context of cross-generational communication. Therefore, this study aimed to classify linguistic styles across generations in social media texts by combining Sentence-BERT (SBERT). FastText-based synonym augmentation in Indonesian, and Support Vector Machine (SVM) as a margin-based classification model that utilizes embedding representations from SBERT. The results showed that synonym augmentation improved model accuracy from 85% to 93%, with a similarity threshold of 0.7 providing the best balance between data diversity and semantic consistency. These findings confirmed that synonym-based augmentation and SBERT semantic adaptation were effective in capturing generational linguistic differences in informal Indonesian. This approach had the potential to be applied in other NLP tasks that required contextual understanding of social language variation, such as sentiment analysis and cross-generational dialect detection.

Analysis and Classification of Capital Assistance Recipients Kediri Trade and Industry Department Using Random Forest

Arika Norma Wahyu Dorroty, Ardiawan Bagus Harisa — Sun, 26 Apr 2026 00:00:00 +0000

Capital assistance provided by the Kediri City Department of Trade and Industry often faces challenges related to the uncertainty of fund distribution, making it difficult to ensure the effectiveness of the assistance itself in improving business revenue. To address this, a prediction-based model is applied to evaluate the factors influencing the success of capital assistance in increasing recipients’ income. This study aims to classify recipients based on business revenue outcomes using the Random Forest algorithm. Furthermore, the model identifies key factors affecting the success of assistance and offers recommendations for optimizing future distribution through feature importance analysis. The results demonstrate that the Random Forest model achieves an accuracy of 75%, highlighting its potential as a reliable tool for predicting the success of capital assistance. The feature importance analysis further reveals that training contributes 49% and business type 43%, emphasizing their crucial role in enhancing the effectiveness of future assistance programs.

Improving Postprandial Glucose Forecasting using Diagnosis-Aware Stacked Learning

Fatma Indriani, Mohammad Reza Faisal, Naufal Said — Sun, 03 May 2026 00:00:00 +0000

Predicting glucose levels after a meal (postprandial glucose) can help anticipate abnormal responses and improve diabetes management. Yet such prediction remains difficult because post-meal glucose depends on multiple interacting factors, including prior glucose trends, meal composition, and recent activity. This study develops machine learning models to forecast short-term post-meal glucose levels using the CGMacros dataset, which combines continuous glucose monitoring (CGM) data from Dexcom and Libre sensors with meal macronutrient annotations and activity measurements. Several feature combinations and regression models were evaluated to identify an optimal representation. Results show that combining baseline glucose statistics with meal composition yields the lowest error across all regressors. Building on this feature configuration, a stacked learning framework was implemented in which a global model provides initial predictions refined by diagnosis-specific CatBoost regressors for Healthy, Pre-diabetes, and Type 2 Diabetes groups. Across 18 configurations spanning two sensors and three horizons (30, 60, 120 minutes), stacking reduced normalized RMSE by 3.5% ± 3.7 on average, with the strongest improvements at 120-minute horizons (mean 5.5%) and for linear global models (up to 13.6% reduction). Gains varied by diagnosis group and sensor type, highlighting the importance of device-aware validation. These results demonstrate that diagnosis-aware stacking enhances both accuracy and robustness, offering a practical foundation for personalized glucose forecasting in digital health systems.

Parameter Efficient Models for Malaria Detection and Classification Using Small-Scale Imbalanced Blood Smear Images

Mon, 27 Apr 2026 00:00:00 +0000

Malaria diagnostic automation faced critical challenges including severe class imbalance with ratios up to 54:1, limited datasets with 200 to 500 images, and computational inefficiency requiring separate model training for each detection-classification combination. This study developed a multi-model framework with shared classification architecture that trained classification models once on ground truth crops and reused them across all detectors. The framework systematically evaluated three YOLO Medium architectures for parasite detection and six CNN architectures for lifecycle and species classification across four complementary malaria datasets totaling 1,544 microscopy images. Detection achieved 70.84% to 96.27% mAP@50 with high recall of 71.05% to 93.12% minimizing missed parasites. Classification demonstrated dataset-dependent model selection with parameter-efficient EfficientNet models containing 5.3M to 9.2M parameters consistently outperforming ResNet variants with up to 44.5M parameters. EfficientNet-B1 achieved 91.51% accuracy on IML Lifecycle and 98.28% on MP-IDB Species, while EfficientNet-B0 achieved 86.45% on multi-patient MD-2019 dataset. ResNet50 achieved 96.13% on severely imbalanced MP-IDB Stages. Focal Loss optimization with alpha of 1.0 and gamma of 1.5 enabled robust minority class performance with F1-scores between 0.44 and 1.00 on ultra-minority classes demonstrating effective imbalance handling. The compact 46-89 MB models enabled practical deployment on resource-constrained hardware.

Revealing Stunting Risk Patterns through Comparative Analysis of Hierarchical and Deep Embedded Clustering

Sun, 03 May 2026 00:00:00 +0000

Stunting remains a significant health issue in Indonesia due to its long-term impact on human resource quality and economic productivity. Despite various intervention programs, disparities in stunting rates between regions remain high, particularly in areas with diverse socioeconomic conditions. This study aims to identify patterns and group regions based on stunting risk levels using two machine learning approaches: Hierarchical Clustering (HC) and Deep Embedded Clustering (DEC). The data used are aggregated data from toddler measurements, including the number of toddlers measured, the number of stunting cases, and the percentage of stunting in the 2020–2024 period. The analysis was conducted by comparing the cluster results from the two methods. The HC method is implemented using an Agglomerative Clustering approach with the Ward linkage criterion, while DEC uses a layered autoencoder architecture optimized through Kullback–Leibler divergence. To assess cluster quality, the study uses the Silhouette Score metric. The results showed that HC produced the highest Silhouette score of 0.5430, while DEC reached 0.4874, with a year-on-year performance trend. These findings indicate that HC excels in clustering stability, while DEC is more adaptive to data complexity and nonlinear patterns. The combination of the two has the potential to support the formulation of more comprehensive, data-driven policies to identify and address stunting-prone areas.

Enhancing CNN Performance for Alzheimer’s Disease Classification Using Genetic Algorithm Optimization

Wildan Arif Maulana, Zainul Abidin, Rahmadwati — Sun, 26 Apr 2026 00:00:00 +0000

The increasing global life expectancy has led to a rapidly growing elderly population, resulting in a higher prevalence of Alzheimer’s disease and a pressing need for effective diagnostic solutions. Despite advances in medical imaging, the early and accurate detection of Alzheimer’s disease remains a major challenge due to subtle differences in brain structures across disease stages. However, the interpretation of MRI images still depends heavily on the abilities of individual medical personnel, which risks introducing subjectivity and potential errors in the diagnostic process. In this context, particularly deep learning, emerges as an effective strategy to overcome these limitations by automating the analysis of medical images and reducing human bias. To address this issue, a custom Convolutional Neural Network (CNN) model was developed from scratch for Alzheimer’s disease classification using brain MRI images. To enhance data diversity and mitigate overfitting, a combination of Albumentations and CutMix data augmentation techniques was applied, yielding an initial classification accuracy of 90%. Model performance was further optimized using a Genetic Algorithm (GA), which efficiently explored the hyperparameter space and identified optimal configurations, boosting classification accuracy to 96%. The optimized model demonstrated robust generalization across all disease categories, confirming the effectiveness of the proposed approach. This research contributes to the development of a more reliable and adaptive deep learning framework for early-stage Alzheimer’s disease detection, offering potential support for clinical diagnostic systems

Performance Analysis of Cluster-based Multi-UAV Routing Protocol in Various Mobility Models using NS-3

Harry Darmawan, Prima Kristalina, Moch. Zen Samsono Hadi — Sun, 26 Apr 2026 00:00:00 +0000

In this paper, the performance of a Cluster-Based Multi-UAV communication system is analyzed as a means to enhance network reliability and coordination in support of Search and Rescue (SAR) operations within disaster-affected area. The proposed approach is intended to address the challenges of maintaining connectivity, ensuring efficient data transmission, and facilitating effective collaboration among UAVs in critical environments. The system is designed with four layers architecture: Base Station (BS), Cluster Head (CH), Clustered Drone (CD), and User Equipment (UE). These layers are modeled and evaluated using Network Simulator 3 (NS-3). The three routing protocols, i.e OLSR, AODV, and DSDV have been evaluated through the three types of UAV mobility: Gauss-Markov, Random Waypoint (RWP). and Reference Point Group Mobility (RPGM). Quality of service parameter for wireless network, such as throughput, packed delivery ratio (PDR), delay, and packet loss has been analysed within several cluster-based UAV schemes. The simulation result shows that cluster-based multi-UAV model with OLSR routing protocol outperforms the best performance in RPGM mobility model with 67.57% average throughput, 87.47% PDR, 86 ms delay, and 12.53% packet loss, better than other routing protocols. OLSR routing protocol indicates the highest consistency with higher throughput and PDR value, smaller delay and packet loss comparing to AODV and DSDV protocols in the small to middle scale of node density. This research contributes in the development of UAV based cluster communication system, especially in efficiency, stabilization and adaptation towards the dynamic environment in the disaster area.

Modeling and Simulation of Heat and Airflow Control System in Fish Smoking Chamber using K-NN

Muhammad Edy Hidayat, Alang Sunding, Umar Muhammad, Irvawansyah — Sun, 26 Apr 2026 00:00:00 +0000

This study presents the modeling and simulation of a heat and airflow control system in a fish smoking chamber using the K-Nearest Neighbors (K-NN) algorithm. Accurate control of temperature and airflow is crucial for ensuring consistent product quality, flavor, texture, and microbial safety in smoked fish. Traditional methods often face challenges in maintaining stable chamber conditions due to nonlinear interactions between heat sources, airflow distribution, and chamber geometry. The research was conducted through a structured methodology consisting of system modeling, K-NN algorithm development, simulation, and performance evaluation. The results show and demonstrate that the K-NN model achieved optimal performance at k = 5, with an overall prediction accuracy of 92.8%. The Root Mean Square Error (RMSE) was recorded at 1.85 °C for temperature prediction and 0.18 m/s for airflow, confirming the model’s robustness. Compared with conventional approaches, K-NN outperformed Linear Regression and achieved higher accuracy with less complexity than Artificial Neural Networks (ANN). The implications of these findings show that predictive modeling enables better process stability, reduces the risk of uneven smoking, and lowers energy consumption. The novelty of this research lies in the dual prediction of heat and airflow, providing a comprehensive framework for smart control in traditional food processing. While the study is limited to simulations, it offers valuable insights for future experimental implementation and integration into intelligent smoking chamber systems.

LITE-BoostTrack: A Hybrid RealTime MultiObject Tracking Architecture for Resource-Constrained Environments

Sun, 26 Apr 2026 00:00:00 +0000

Multi object tracking (MOT) is a crucial component of modern computer vision applications, ranging from intelligent surveillance to autonomous vehicles. The primary challenge in MOT lies in maintaining identity consistency under conditions of high density and frequent occlusion, while also ensuring computational efficiency for real time deployment on resource constrained devices. This paper introduces LITE BoostTrack, a hybrid architecture that combines the confidence scaling based association mechanism of BoostTrack with the lightweight feature extraction strategy of the Lightweight Integrated Tracking and Embedding (LITE) framework. By leveraging internal features from the YOLOv8 detector without relying on an external Re Identification module, the proposed approach reduces computational burden while preserving robustness in identity association. Experiments were conducted on the MOT20 benchmark using standard evaluation metrics, namely HOTA, MOTA, IDF1, IDSW, and FPS, to comprehensively assess both tracking accuracy and runtime efficiency. The results demonstrate that LITE BoostTrack achieves competitive accuracy, with a HOTA of 27.32 and an IDF1 of 37.49, which are nearly equivalent to the original BoostTrack. At the same time, it delivers a substantial improvement in runtime efficiency, reaching 13.23 FPS, almost twice the speed of standard BoostTrack. These findings confirm that efficiency optimization in MOT can be achieved through architectural reengineering that exploits detector internal features without the need for additional deep modules. LITE BoostTrack therefore represents a balanced and practical solution that combines accuracy with efficiency, making it well suited for real time applications in edge computing and resource constrained environments.

Distributed Secondary Control with Consensus-Based Adaptive Droop and Voltage Observer for DC Microgrids

Khusnul Hidayat, Arif Nur Afandi — Sun, 03 May 2026 00:00:00 +0000

This paper proposes a fully distributed secondary control scheme for a low-voltage DC microgrid with ring topology. The main objectives are to restore the common bus voltage to its nominal reference and to achieve accurate proportional current sharing among distributed generator units in the presence of non-uniform line resistances and mixed load conditions. The proposed secondary layer integrates a consensus-based adaptive droop controller and a consensus-based voltage observer. The adaptive droop mechanism dynamically adjusts the virtual impedance of each converter using neighbor-to-neighbor current information to reduce current-sharing errors, while the voltage observer provides a distributed estimate of the average bus voltage to compensate for droop-induced voltage deviations. The effectiveness of the proposed method is validated through simulation on a ring-configured DC microgrid consisting of four converters and five buses. A comparative study demonstrates that conventional droop control improves current sharing but introduces significant steady-state voltage deviation. By contrast, the proposed integrated approach achieves nearly zero current-sharing error while maintaining the DC bus voltage close to its reference value. The dynamic performance is further evaluated under both resistive-load and constant-power-load variations. The results show that the controller ensures fast voltage restoration, accurate proportional current sharing, and stable operation without sustained oscillations, even under nonlinear constant-power-load conditions. These findings indicate that the proposed distributed secondary control strategy provides robust voltage regulation and precise current sharing for ring-type DC microgrids.

Performance Evaluation of Motion Estimation and Compensation Algorithms in SNR Scalable Video Encoding

Sun, 26 Apr 2026 00:00:00 +0000

Motion estimation is the sequential determination of the direction of motion of an object in a video. The movement of an object is denoted by the term motion vector. Between the current and reference frames, motion vectors can signify shift points. The SAD (Sum of Absolute Different) block matching technique is fundamentally dependent on the assessment of an object's motion. In this study proposes a hybrid approach that integrates the Three-Step Search (TSS) and Full Search (FS) algorithms. This integration aims to design a block matching algorithm that is applied to video encoding using signal-to-noise ratio (SNR) scalability. From this design, we hope to obtain the performance and evaluate the motion estimation process utilizes both the TSS and FS algorithms for performance comparison on SNR scalability video encoding to obtain video frame quality in relation to bit rate and PSNR, based on the average comparison of the two algorithms. Based on the experimental results, the FS algorithm achieved a total BD-PSNR of 0.22 dB with an efficiency rate of 12.45%, whereas the TSS algorithm achieved a total BD-PSNR of 0.18 dB and an efficiency rate of 7.6%. Therefore, the FS algorithm demonstrates superior performance compared to the proposed TSS algorithm in video transmission with SNR scalability.

Integrating Tabular Data and Textual Representations for Clinical Risk Prediction Using Machine Learning and Large Language Models

Sun, 03 May 2026 00:00:00 +0000

Global health is currently facing serious challenges due to the increasing number of chronic disease patients, such as those with heart failure, diabetes, and cancer. This issue arises from the limitations of electronic health record (EHR) systems, which are not yet fully capable of ensuring accurate clinical diagnoses because of potential data input errors and delays in symptom identification by medical personnel. In response to this issue, this paper focuses on the integration of medical tabular data with a classification approach based on classical machine learning (ML) and large language models (LLM) to improve the accuracy of patient diagnosis predictions. This paper aims to develop and compare the performance of various ML models, such as XGBoost, SVM, and logistic regression, as well as LLM models like Gemini, LLaMA, and Qwen in fine-tuning, few-shot, and zero-shot scenarios. The paper results show that the combination of Gemini and the few-shot approach (250 shots) achieved the highest accuracy of up to 99.8% in predicting heart failure risk. The main finding of this study is that the narrative text representation of tabular data processed with LLM significantly enhances contextual understanding and classification accuracy, making this approach highly potent for application in AI-based clinical decision-making.

Imitation Learning to Accelerate Training Process of Multi-Agent Reinforcement Learning in 2v2 Pong Game

Marvin Yonathan Hadiyanto, Budi Harsono, Indra Karnadi, Ivan Tanra — Sun, 03 May 2026 00:00:00 +0000

Training multi-agent reinforcement learning (MARL) systems often requires a significant amount of time due to sample inefficiency, particularly where agents need to do a considerable amount of exploration in a complex environment and coordination among multiple entities. This study proposes the use of imitation learning to accelerate the MARL training process in a 2v2 pong game by learning from demonstrations in 1v1 pong game to shape the initial policy without undergoing inefficient exploration procedure. We use deep Q-network (DQN) with centralized training with decentralized execution (CTDE) to observe the difference of performance between pretrained and untrained agents in 2v2 pong game. Experimental results show that learning from demonstration in 1v1 setting significantly improved reward accumulation and game scores of pretrained agent in 2v2 pong game. The improvement peaks at 700 learning steps of demonstration and diminishes at the larger learning steps due to excessive memorization of the demonstration gameplay. This work demonstrates that imitation learning from demonstrations can be used to reduce a prolonged training process in MARL, offering a viable solution especially when data collecting, computational resources, and training are the severely constrained.