AI-Powered Search and Navigation

Introduction

As part of my coursework, I designed intelligent agent systems for dynamic, uncertain environments, progressively building and analyzing increasingly complex AI models. These projects spanned search algorithms, probabilistic reasoning, and machine learning-based predictions.

Fire Extinguisher Bot (Project 1)

Objective

Build bots to navigate a spaceship under fire hazards and press an emergency button before getting trapped.

Skills & Tools Used:

1. Graph Search (A*, Uniform Cost Search)
2. Heuristic Design (Manhattan, Euclidean, Chebyshev distances)
3. Risk-aware pathfinding using dynamic fire spread prediction
4. Success rate analysis across varying environmental flammability (parameter q)

Highlights:

1. Created, tuned and tested 10+ bot strategies.
2. Final bot (Bot 9) used a cost function balancing distance, immediate fire risk, and future spread risk.
3. Achieved the highest average success rate (74.4%) across simulations.

---

Localization and Target Tracking (Project 2)

Objective

Build bots that localize themselves in an unknown grid and capture a hidden "space rat" using partial noisy observations.

Skills & Tools Used:

1. Bayesian Networks for probabilistic localization
2. Utility-based decision-making (entropy reduction)
3. Dynamic re-planning with BFS (Breadth-First Search)
4. Handling both stationary and randomly moving targets

Highlights:

1. Designed baseline and two improved bots.
2. Incorporated a transition model for moving targets, significantly improving tracking performance.
3. Conducted 100+ simulation experiments per setup; statistical tests validated improvements.

---

Predicting Capture Time with Neural Networks (Project 3)

Objective

Predict the remaining time steps needed for a bot to catch a space rat, using probabilistic state representations.

Skills & Tools Used:

1. Convolutional Neural Networks (CNNs)
2. Regression tasks (MSE, MAE loss functions)
3. Dataset generation from probabilistic simulations (~27,000 samples)
4. Model optimization (batch normalization, dropout, residual connections)

Highlights:

1. Designed a lightweight CNN processing 2×30×30 probability matrices.
2. Achieved low test losses (MSE: 974.85; MAE: 17.71) relative to task variance.
3. Showcased effective generalization through careful regularization strategies.