A public project archive for optimization coursework on Markov decision processes, tabular reinforcement learning, and image-based Deep Q-Network training in a custom MiniGrid maze.
Project3_solution_최은정안상호.pdf
The course covered unconstrained nonlinear optimization, optimization algorithms, knapsack problems, artificial intelligence and machine learning fundamentals, Markov decision processes, and reinforcement learning. Python and PyTorch were used to connect the mathematical concepts with implementation-based projects.
We designed a deterministic 8 × 8 MiniGrid maze with a 6 × 6 traversable interior, fixed start and goal states, dead ends, and narrow turns. The shortest feasible route requires 15 steps. The project progressed from tabular reinforcement-learning baselines to a CNN-based DQN agent operating on a restricted local observation.