Momin Haider - MS CS Student at UCSB

Projects:

Built a detailed browser-based platformer that can learn to play itself by training with Proximal Policy Optimization (PPO)
Implemented local storage of network weights and biases, allowing anyone with a browser and internet access to train the reinforcement learning agent
Utilized TensorFlow.js to store and train actor/critic parameters as well as to choose actions based on hand-engineered input feature vectors
Implemented custom garbage collection routines to dispose tensors and prevent memory leaks within the browser

Trained an agent using Twin Delayed DDPG (TD3) to solve the Bipedal Walker challenge from OpenAI
Used PyTorch to store parameters that define actor and critic networks as well as to perform gradient descent
Allowed the training routine to make use of CUDA if possible to perform backpropagation on an external GPU
Wrote a supplemental document explaining the TD3 algorithm implemented assuming initial knowledge of Deep Q-Learning and the Bellman equation

Developed an extensive animation library in Python that interfaces with Blender
Structured a rendering interface that interacts with FFMPEG via subprocess calls, allowing flexible frame-by-frame manipulation
Made use of quaternions, rotation matrices, numerical integration methods, and Bézier interpolation, among other necessary math concepts

Research Assistant (Reinforcement Learning): UC Santa Barbara (Apr 2023 - Present)
Teaching Assistant (Fundamentals of Database Systems): UC Santa Barbara (Sep 2023 - Present)
Teaching Assistant (Intro to Computer Science): UC Santa Barbara (Sep 2022 - Dec 2022)
Founding Software Engineer: PeepPhysics.com (Jan 2020 - Oct 2021)
Teaching Assistant (Aircraft Stability): UF (Aug 2019 - Dec 2019)
Research Assistant (Biophysics): UF (Feb 2017 - Jan 2019)

El Porto (2019)