Publications
My research contributions in reinforcement learning, social robotics, and neural network architectures.
Featured Publication
Cosine-Gated LSTM for Sequential Learning
Demonstration of Cosine-Gated LSTM performance in FallingBallEnv environment
Novel Cosine-Gated LSTM (CGLSTM) architecture that uses cosine gating mechanisms for improved sequential learning. The architecture dynamically weighs inputs versus hidden states based on cosine similarities, outperforming LSTM, GRU, and Transformer models in benchmark tasks.
View Full Abstract & Details
Abstract
Introduced a novel Cosine-Gated LSTM (CGLSTM) architecture that uses cosine gating mechanisms for improved sequential learning. The architecture dynamically weighs inputs versus hidden states based on cosine similarities, outperforming LSTM, GRU, and Transformer models in benchmark tasks including FallingBallEnv and SocNavGym environments.
Key Contributions
- Novel cosine gating mechanism for LSTM architecture
- Dynamic input weighting based on cosine similarity
- Superior performance compared to baseline architectures
- Efficient parameter usage with improved long-term dependency handling
All Publications
Filter by:
Predictive World Models for Social Navigation
Predictive World Models in action: Social navigation with 2StepAhead and MASPM
Developed predictive world models enabling autonomous systems to navigate safely in social environments. The work introduces 2StepAhead and Multi-Action State Predictive Models (MASPM) that enable multi-step ahead predictions and comprehensive action evaluation.
View Full Abstract & Details
Abstract
Developed predictive world models enabling autonomous systems to navigate safely in social environments. The work introduces 2StepAhead and Multi-Action State Predictive Models (MASPM) that enable multi-step ahead predictions and comprehensive action evaluation, significantly improving navigation performance in dynamic social environments.
Key Contributions
- 2StepAhead model for multi-step ahead predictions
- MASPM (Multi-Action State Predictive Model) for comprehensive action evaluation
- Hybrid model combining both approaches
- Superior performance compared to baseline methods
Enhancing Robot Social Navigation with Reinforcement Learning and Advanced Predictive Models
Adaptive Prediction Horizons in action: LunarLander-v2 environment
Comprehensive research integrating reinforcement learning with advanced predictive models for social robot navigation. The research introduces three key contributions: Predictive World Models (2StepAhead and MASPM), Cosine-Gated LSTM (CGLSTM) architecture, and Adaptive Prediction Horizons.
View Full Abstract & Details
Abstract
This PhD thesis presents comprehensive research integrating reinforcement learning with advanced predictive models for social robot navigation. The research introduces three key contributions that collectively advance the field: Predictive World Models (2StepAhead and MASPM), Cosine-Gated LSTM (CGLSTM) architecture, and Adaptive Prediction Horizons. The work demonstrates superior performance in simulation environments including LunarLander-v2 and SocNavGym, achieving higher success rates and more efficient navigation compared to baseline methods.
Key Contributions
1. Predictive World Models
- 2StepAhead: Multi-step ahead prediction system
- MASPM: Multi-Action State Predictive Model for comprehensive action evaluation
- Hybrid approach combining both models for optimal performance
2. Cosine-Gated LSTM Architecture
- Novel architecture using cosine gating mechanisms
- Dynamic input weighting based on cosine similarity
- Superior sequential learning performance
3. Adaptive Prediction Horizons
- Entropy-driven adaptive prediction horizon systems
- Dynamic adjustment based on environmental complexity
- Improved performance in diverse RL environments
Results
- Successfully defended PhD thesis with excellent feedback
- Published 2 conference papers (UKCI 2023, IEEE PRML 2024)
- Superior performance in simulation environments
- Higher success rates than baseline methods
- More efficient navigation algorithms
Citation Tools
If you use my work in your research, please cite appropriately. Citation tools with BibTeX, APA, MLA, and Chicago formats are available for each publication above.
For questions about my publications, collaboration opportunities, or accessing materials, please contact me.