Shiladitya Biswas

A simple Matrix library to perform Matrix operations namely Matrix Multiplication (using Stressan’s Algorithm) and Transpose.

We applied the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to a multi-agent reinforcement learning scenario involving a four-lane traffic intersection. The performance of MADDPG was compared with the Proximal Policy Optimization (PPO) algorithm to evaluate their effectiveness. To enhance the stability of MADDPG, we incorporated prioritized experience replay (PER) and conducted a detailed analysis of how factors such as batch size, multi-CPU training, and experience collection influence agent performance. We used CARLA for this project. This was a course project for ECE276C at UC San Diego.

Implemented and compared the performance of value and Policy iteration-based controller in terms of convergence time and state space resolution on an Inverted pendulum balancing controller in Python. Formulated a Markov Decision Process (MDP) for the system by defining a discreet state space over the angular velocity, angular postion and control input of the pendulum. A transtion matrix is built over the 3D state-space of the system by using Euler discretization for state propgation and brownian noise to model the uncertanities in the system. Finally, a cost function is defined over the state-space which is ultimately used to solve the infinite-horizon discounted optimal control problem, thereby generating the optimal control policies. This was a course project for ECE276B at UC San Diego, hence the code cannot me made public.

Implemented and compared the performance of A* and Rapidly expanding Random Tree (RRT*) Algorithm in terms of planning time and path length on multiple 3D environments with obstacles in Python. This was a course project for ECE276B at UC San Diego, hence the code cannot me made public.

The goal here was to figure out a set of control policy that enables the robot (red triangle) to first fetch the key, open the door with the key and then reach the goal location. To achive this, I formulated the problem as a Markov Decision Process (MDP) in which each cell on the 2D grid has a reward associted to it. I then implemented Dynamic Programming to generate the optimum control policy. This was a course project for ECE276B at UC San Diego, hence the code cannot me made public.

Implemented an Extended Kalman Filter based Visual-Inertial SLAM algorithm to track the three dimensional position and orientation of a Car. Used data from gyroscope, accelerometer, and camera measurements. Validated the algorithm on real world KITTI Dataset. This was a course project for ECE276A at UC San Diego, hence the code cannot me made public.

Evaluated the Particle filter using real-world odometry data, indoor 2D laser scans, and RGBD measurements captured by THOR, a humanoid robot equipped with LiDAR and Kinect v2 sensors. The Particle filter successfully demonstrates its effectivity in capturing highly non-linear motion and measurement models. I used Sample Importance Sampling for resampling the particles inorder to address the particle depletion phenomenon. This was a course project for ECE276A at UC San Diego, hence the code cannot me made public.

Implemented Gaussian Discriminant Analysis (GDA)- a supervised learning algorithm, to classifying Red pixles from non-red pixels, and thereby detected stop signs.This was a course project for ECE276A at UC San Diego, hence the code cannot me made public.

Applied Principal Component Analysis (PCA) to a collection of face images to generate a set of basis features. Projected face images onto a feature space (”face space”) optimized to encode the variation among known face images. Defined the face space using ’eigenfaces’, which are the eigenvectors derived from the set of faces and represent patterns of variation rather than isolated facial features like eyes, ears, or noses.

This project marked one of my earliest endeavors into creating a camera-based self-driving car. The car navigated autonomously by following a white path on a black background. A smartphone was utilized to transmit live video feeds to a laptop, where the images underwent pyramidal reduction processing and were subsequently fed into a neural network. The neural network’s output layer comprised of four nodes—Left, Right, Forward, and Backward—responsible for decision-making. These decisions were then relayed back to the car via Bluetooth and executed by an Arduino Uno board, enabling the vehicle to maneuver along the road. This project served as a proof of concept for my UG thesis project “A Cloud based End to End Self Driving Car Prototype”. The training dataset that was collected for this project can br found here.