Selected work

Representative Papers

An Agency-Transferring Model-Free Policy Enhancement Technique

Jun 08, 2026

This work addresses the inefficiency of existing reinforcement learning methods that often disregard available suboptimal baseline policies, resulting in high training costs and low task success rates. We propose a model-free policy augmentation framework that leverages a dynamic arbitration mechanism: during early training, control is delegated to a functional baseline policy to ensure goal reachability, and is gradually transferred to a learnable policy, ultimately yielding a high-performance policy independent of the baseline. We formally define functional baselines for the first time and integrate probabilistic reachability analysis to design the transfer mechanism, providing theoretical guarantees on the lower bound of goal achievement probability for the final policy. Experiments on continuous control benchmarks demonstrate that our method achieves competitive or superior returns compared to state-of-the-art approaches while consistently maintaining the highest goal success rate throughout both training and standalone deployment.

0 citationsRead paper

Virtual Holonomic Constraints in Motion Planning: Revisiting Feasibility and Limitations

May 12, 2025

This work addresses the limited applicability of Virtual Holonomic Constraints (VHCs) in motion planning for single-degree-of-freedom underactuated systems—such as the Planar Vertical Takeoff and Landing (PVTOL) aircraft—where conventional VHC definitions are overly restrictive, excluding many feasible periodic trajectories. To overcome this limitation, we propose a reformulated VHC framework that relaxes geometric and realizability requirements on the constraint manifold. Leveraging phase-trajectory analysis, analytic function modeling, and nonlinear feedback control synthesis, we design a controller ensuring asymptotic orbital stability. Theoretical analysis and experimental validation demonstrate that a class of analytic periodic solutions previously inadmissible under classical VHC theory can now be precisely characterized and robustly stabilized within the new framework. This advances VHC theory by significantly broadening its scope and practical utility in underactuated system motion planning.

0 citationsRead paper

Deep Learning Models Meet Financial Data Modalities

Apr 18, 2025

Modeling multi-source heterogeneous financial data—such as candlestick charts, order flow, trading volume, limit order book (LOB) snapshots, and news—poses significant challenges due to their disparate structures and temporal dynamics. To address this, we propose a unified multimodal deep learning framework. Our key innovation is the first-ever encoding of LOB snapshot sequences as multi-channel images, coupled with a dedicated embedding scheme that enables visual representation learning for structured time-series data. We further design a hybrid CNN–RNN architecture to jointly model cross-modal features, support inter-modal alignment, and learn dynamic modality-specific weights. Evaluated on high-frequency trading strategy tasks, our method achieves state-of-the-art performance: it improves price direction prediction accuracy by +3.2% and portfolio Sharpe ratio by +0.41. This work establishes a scalable, multimodal deep learning paradigm for financial time-series modeling.

0 citationsRead paper

Multimodal Stock Price Prediction: A Case Study of the Russian Securities Market

Mar 05, 2025Program systems theory and applications

This study addresses the challenge of stock price forecasting in the Russian equity market by proposing a multimodal approach that jointly models candlestick time-series data and Russian-language financial news texts. To capture news semantics in the Russian financial domain, we systematically integrate RuBERT and Vikhr-Qwen2.5-0.5b-Instruct—constituting the first such deployment—and couple them with an LSTM for price sequence modeling; cross-modal synergy is achieved via feature concatenation and learnable weighted fusion. Experimental results demonstrate that incorporating news modality reduces MAPE by 55% and improves directional prediction accuracy. Furthermore, we release the first open-source multimodal benchmark dataset for Russian finance, comprising candlestick sequences of 176 Russian-listed stocks and 79,000 manually annotated Russian financial news articles—filling a critical gap in multilingual financial multimodal resources.

0 citationsRead paper

Recent publications

Latest Papers

An Agency-Transferring Model-Free Policy Enhancement Technique

Jun 08, 2026

This work addresses the inefficiency of existing reinforcement learning methods that often disregard available suboptimal baseline policies, resulting in high training costs and low task success rates. We propose a model-free policy augmentation framework that leverages a dynamic arbitration mechanism: during early training, control is delegated to a functional baseline policy to ensure goal reachability, and is gradually transferred to a learnable policy, ultimately yielding a high-performance policy independent of the baseline. We formally define functional baselines for the first time and integrate probabilistic reachability analysis to design the transfer mechanism, providing theoretical guarantees on the lower bound of goal achievement probability for the final policy. Experiments on continuous control benchmarks demonstrate that our method achieves competitive or superior returns compared to state-of-the-art approaches while consistently maintaining the highest goal success rate throughout both training and standalone deployment.

0 citationsRead paper

Virtual Holonomic Constraints in Motion Planning: Revisiting Feasibility and Limitations

May 12, 2025

This work addresses the limited applicability of Virtual Holonomic Constraints (VHCs) in motion planning for single-degree-of-freedom underactuated systems—such as the Planar Vertical Takeoff and Landing (PVTOL) aircraft—where conventional VHC definitions are overly restrictive, excluding many feasible periodic trajectories. To overcome this limitation, we propose a reformulated VHC framework that relaxes geometric and realizability requirements on the constraint manifold. Leveraging phase-trajectory analysis, analytic function modeling, and nonlinear feedback control synthesis, we design a controller ensuring asymptotic orbital stability. Theoretical analysis and experimental validation demonstrate that a class of analytic periodic solutions previously inadmissible under classical VHC theory can now be precisely characterized and robustly stabilized within the new framework. This advances VHC theory by significantly broadening its scope and practical utility in underactuated system motion planning.

0 citationsRead paper

Deep Learning Models Meet Financial Data Modalities

Apr 18, 2025

Modeling multi-source heterogeneous financial data—such as candlestick charts, order flow, trading volume, limit order book (LOB) snapshots, and news—poses significant challenges due to their disparate structures and temporal dynamics. To address this, we propose a unified multimodal deep learning framework. Our key innovation is the first-ever encoding of LOB snapshot sequences as multi-channel images, coupled with a dedicated embedding scheme that enables visual representation learning for structured time-series data. We further design a hybrid CNN–RNN architecture to jointly model cross-modal features, support inter-modal alignment, and learn dynamic modality-specific weights. Evaluated on high-frequency trading strategy tasks, our method achieves state-of-the-art performance: it improves price direction prediction accuracy by +3.2% and portfolio Sharpe ratio by +0.41. This work establishes a scalable, multimodal deep learning paradigm for financial time-series modeling.

0 citationsRead paper

Multimodal Stock Price Prediction: A Case Study of the Russian Securities Market

Mar 05, 2025Program systems theory and applications

This study addresses the challenge of stock price forecasting in the Russian equity market by proposing a multimodal approach that jointly models candlestick time-series data and Russian-language financial news texts. To capture news semantics in the Russian financial domain, we systematically integrate RuBERT and Vikhr-Qwen2.5-0.5b-Instruct—constituting the first such deployment—and couple them with an LSTM for price sequence modeling; cross-modal synergy is achieved via feature concatenation and learnable weighted fusion. Experimental results demonstrate that incorporating news modality reduces MAPE by 55% and improves directional prediction accuracy. Furthermore, we release the first open-source multimodal benchmark dataset for Russian finance, comprising candlestick sequences of 176 Russian-listed stocks and 79,000 manually annotated Russian financial news articles—filling a critical gap in multilingual financial multimodal resources.

0 citationsRead paper