Yash Guleria Research Fellow, ATMRI, NTU

About Me Interests Publications

Hey there!

Professional Me

Hi, I am Yash. I am a research fellow at the Air Traffic Management Research Institute (ATMRI), Nanyang Technological University (NTU), Singapore. Here, I work at the intersection of machine learning, optimization, and air traffic control, to address the near future air traffic demands.

Currently, I am leading a team of 3 scientists to utilize large language models to generate safety-critical scenarios in the en-route phase of the aircraft. The primary motivation for this work stems from the scarcity of conflict scenarios in the historical data and the complexity and iterations involved in creating such scenarios, and the difficulty in customization and interactive enhancement of the traffic scenarios using traditional techniques. I received my Ph.D. in 2024 , with a focus of devloping machine learning models for air traffic conflict resolution for increased acceptance by the air traffic controllers.

Research Interest

  • Large Language Models (LLMs)
  • Deep Learning
  • Reinforcement Learning
  • Data-centric AI
  • Applied AI
  • Intelligent Transportation systems

    • Publications and projects

    For a complete list of publications, please visit my Google Scholar profile. Google Scholar.
    A few recent projects are listed below:

    Text2Traffic: Retrieval Enhanced In-Context Learning For Complex Air Traffic Scenario Generation

  • Developed a RAG-enabled LLM approach to generate complex air traffic scenarios.
  • Compared retrieval capabilities of models such as Cohere's Command r, Llama3.1-8b-Instruct, and GPT3.5-Turbo.
  • The methodology enables quick and customizable air traffic scenario generation based on natural language inputs from the user.
    • Conformal automation diagram

    Text2Traffic Methodology: Based on user input, Text2Traffic leverages Retrieval-Augmented Generation (RAG) to extract relevant sector information, including airways, waypoints, and aircraft types, from a domain-specific document, enabling the generation of the requested air traffic scenario.

    Tools and technologies used
    Retrieval Augmented Generation (RAG) LLMs (GPT3.5-Turbo, Llama3.1-8b-Instruct) Langchain Syntactic and semantic performance evaluation Maximal Marginal relevance (MMR)

    Towards conformal automation in air traffic control: Learning conflict resolution strategies through behavior cloning

  • Developed a novel methodology incorporating supervised machine learning, to predict air traffic controllers’ conflict resolution strategies.
  • Developed an experimental interface in collaboration with EUROCONTROL, to conduct human-in-loop experiments with expert air traffic controllers from Singapore and France.
  • Performed human-in-loop validation experiments to test the acceptability of machine-learning prediction for air traffic conflict resolution.
  • Developed a reinforcement learning-based model to incorporate air traffic controller’s knowledge into an agent capable of performing ATCO-like conflict resolution
    • Conformal automation diagram

    End-to-end conflict resolution pipeline using ATCO-conformal ML models.

    Tools and technologies used
    Supervised machine learning (Random forest, XGBoost, SVMs) Human-in-the-loop experiments Data analytics and Visualization Model robustness analysis

    An Agent-Based Approach for Air Traffic Conflict Resolution in a Flow-Centric Airspace

  • Conceptualized and investigated air traffic conflict resolution in flow-centric operations, where the traffic was modeled as intersecting flows.
  • Developed a novel agent-based model to resolve conflicts in flow-centric airspace, where a deep reinforcement learning-based agent was responsible for resolving flow-based air traffic conflicts in enroute sectors.
    • Flow-centric conflict resolution

    A concept diagram for the interaction between the agent and the learning environment. Scenarios involving interflow and intra-flow conflict are generated and the vector representation of the extracted features is used by the agent to propose an action based on the learned policy, thereby reaching a new state and receiving a certain reward. The updated actions pass into a self-stabilizing algorithm which ensures intra-flow safe separation and outputs the updated location and speed of the aircraft in both flows.

    Tools and technologies used
    Deep reinforcement learning (PPO, DQN) Agent-based modeling

    Continuous descent flight operations in Singapore Changi airport.

  • Worked with a team of 4 scientists on the development of an ML approach for continuous descent prediction for aircraft arriving at Singapore Changi airport.
  • Liaised with THALES AirLab for integration of the ML methodologies into an air traffic control digital twin for evaluation by air traffic controllers.
    • Tools and technologies used
    Random forest regression, feature engineering

    Conferences / Seminars / Workshops

    • Collaborated with the European Union Aviation Safety Agency (EASA) for the organization of the EU - Asia Pacific Artificial Intelligence in Aviation Workshop, hosted by ATMRI, 21st – 22nd November 2024.
    • Organization committee member of the 11th International Conference on Research in Air Transportation (ICRAT) 2024, Singapore (1st - 4th July 2024).
    • Conducted a 3-day workshop on the application of reinforcement learning in air transportation at the European Union Aviation Safety Agency (EASA) headquarters, Cologne, Germany from 11th to 13th September 2023.