Кафедра мультимедійних систем

Permanent URI for this collection

https://ekmair.ukma.edu.ua/handle/123456789/12253

Browse

Browsing Кафедра мультимедійних систем by Author "Kuzmenko, Dmytro"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Energy Conservation for Autonomous Agents Using Reinforcement Learning
    (2025) Beimuk, Volodymyr; Kuzmenko, Dmytro
    Reinforcement learning (RL) has shown strong potential in autonomous racing for its adaptability to complex and dynamic driving environments. However, most research prioritizes performance metrics such as speed and lap time. Limited consideration is given to improving energy efficiency, despite its increasing importance in sustainable autonomous systems. This work investigates the capacity of RL agents to develop multi-objective driving strategies that balance lap time and fuel consumption by incorporating a fuel usage penalty into the reward function. To simulate realistic uncertainty, fuel usage is excluded from the observation space, forcing the agent to infer fuel consumption indirectly. Experiments are conducted using the Soft Actor-Critic algorithm in a high-fidelity racing simulator, Assetto Corsa, across multiple configurations of vehicles and tracks. We compare various penalty strengths against the non-penalized agent and evaluate fuel consumption, lap time, acceleration and braking profiles, gear usage, engine RPM, and steering behavior. Results show that mild to moderate penalties lead to significant fuel savings with minimal or no loss in lap time. Our findings highlight the viability of reward shaping for multi-objective optimization in autonomous racing and contribute to broader efforts in energy-aware RL for control tasks. Results and supplementary material are available on our project website.
  • Thumbnail Image
    Item
    Hybrid AI Model for Financial Market Prediction
    (2025) Voitishyn, Mykyta; Kuzmenko, Dmytro
    Financial time series modeling is increasingly complex due to volatility, unexpected breakouts, and the impact of external factors, such as macroeconomic indicators, investor sentiment, company fundamentals, and extreme shocks, like geopolitical events or market manipulations. This paper introduces a hybrid artificial intelligence framework that integrates traditional statistical methods, machine learning models, and Bayesian neural networks (BNNs) to improve predictive performance and uncertainty quantification in financial forecasting. The model leverages a variety of engineered features, including rolling statistics, technical indicators, anomaly scores, interpolated macroeconomic data, and transformer-based sentiment scores. A complete ablation study compares various architectures, including ARIMA, SARIMA, MLR, SNN, and BNN, across multiple prediction windows (1, 3, 5 days) and feature combinations. Results show that while linear models yield the lowest MSE for short-term predictions, they fail to capture non-linear dependencies and uncertainty. In contrast, BNNs offer more reliable mid-term predictions by estimating predictive distributions. The best BNN configuration (Normal distribution, constant variation, TanH activation, 1 hidden layer) achieved an MSE of 0.00022, confirming the advantage of uncertainty-adjusted modeling. Sentiment analysis and anomaly detection were especially impactful when combined with macroeconomic indicators, improving signal reliability and behavioral insight. Our findings highlight the importance of integrating diverse data sources and accounting for predictive uncertainty in financial applications. Additionally, the experiments revealed that compact network architectures often outperform deeper ones when paired with engineered features. All experiments were systematically tracked to ensure reproducibility and facilitate future model benchmarking.
  • Thumbnail Image
    Item
    Knowledge Transfer in Model-Based Reinforcement Learning Agents for Efficient Multi-Task Learning
    (2025) Kuzmenko, Dmytro; Shvai, Nadiya
    We propose an efficient knowledge transfer approach for modelbased reinforcement learning, addressing the challenge of deploying large world models in resource-constrained environments. Our method distills a high-capacity multi-task agent (317M parameters) into a compact 1M parameter model, achieving state-of-the-art performance on the MT30 benchmark with a normalized score of 28.45, a substantial improvement over the original 1M parameter model’s score of 18.93. This demonstrates the ability of our distillation technique to consolidate complex multi-task knowledge effectively. Additionally, we apply FP16 post-training quantization, reducing the model size by 50% while maintaining performance. Our work bridges the gap between the power of large models and practical deployment constraints, offering a scalable solution for efficient and accessible multi-task reinforcement learning in robotics and other resource-limited domains.