Selected Publications

Towards Better Exploitation-and-Exploration Balancing in Time Series Explanation via Reinforcement Learning Fuyuan Lyu, Melissa Mozifian TODO Under Submission

Robust Reinforcement Learning Objectives for Sequential Recommender Systems Melissa Mozifian, Tristan Sylvain, Dave Evans, Lili Meng arXiv Code / Bibtex

Learning Successor Feature Representations to Train Robust Policies for Multi-task Learning Melissa Mozifian, Dieter Fox, David Meger, Fabio Ramos, Animesh Garg NeurIPS Workshop on Deep Reinforcement Learning (NeurIPS), 2022

Intervention Design for Effective Sim2Real Transfer Melissa Mozifian, Amy Zhang, Joelle Pineau, and David Meger arXiv Code / Video / Bibtex

Shaping Rewards for Combined Reinforcement and Imitation Learning using Generative Models Yuchen Wu, Melissa Mozifian, Florian Shkurti International Conference on Computer Vision on Robotics and Automation (ICRA), 2021 Code / Video / Bibtex

	Learning Domain Randomization Distributions for Training Robust Locomotion Policies Melissa Mozifian*, Juan Camilo Gamboa Higuera*, David Meger, Gregory Dudek International Conference on Intelligent Robots and Systems (IROS), 2020 Multi-Task and Lifelong Reinforcement Learning Workshop (ICML), 2019 Code / Bibtex
	Joint 3D Proposal Generation and Object Detection from View Aggregation Jason Ku, Melissa Mozifian, Ali Harakeh, Jungwook Lee, Steven Waslander International Conference on Intelligent Robots and Systems (IROS), 2018 Code / Kitti Benchmark (AVOD) / Paper Video / More Video/ Bibtex
	Deep Learning a Quadrotor Dynamic Model for Multi-Step Prediction Nima Mohajerin, Melissa Mozifian and Steven Waslander International Conference on Computer Vision on Robotics and Automation (ICRA), 2018 Quadrotor Dataset / Bibtex

Past Projects

• Real-time 3D Object Detection for Autonomous Driving (2018)

  Master thesis.

  AVOD-SSD Code

• Autonomoose (2017)

  Worked on deploying our 3D object detector, AVOD, with ROS integration on our self-driving car. See video

• Neural Network based EKF Localization (2017)

  Object-based localization method with neural networks using YOLO (real-time object detector) to localize objects in the camera frame. Using these detected objects as features, an Extended Kalman Filter was used to estimate the robot pose. See video

• Modelling Quadrotor Dynamics Using Neural Networks (2017)

  Using deep recurrent neural network architecture to model quadrotor dynamics with autoencoders as a pre-training technique.

• Kaggle Seizure Prediction Challenge (2016)

  Electrical brain activity (EEG) based seizure forecasting systems using machine learning methods. See the challenge

• Affective Mirror (BSc Dissertation - 2014)

  The goal of this project was to develop a system with the ability to detect emotions such as happiness, sadness, surprise and excitement by analysing the facial cues using SVMs. See video

• Webcam-based Pulse Detector (2013)

  This application uses OpenCV to detect the user's face to isolate the forehead region. Data is collected from this region over time to estimate the user's heart rate by measuring average optical intensity in the forehead location. Physiological data can be estimated this way due to the optical absorption characteristics of (oxy-) haemoglobin See Plethysmographic imaging

• Idris, A Dependently Typed Functional Programming Language (2012)

  I contributed to Idris, a general-purpose functional programming language with dependent types. I implemented a Secure SQL Query system Using Dependent Types.