Chen Yijia, Victoria Chong, Sim Shang Hong
In this project, we use various AI techniques - from classical to more advanced deep learning style - to help agents navigate mazes. We compared to performance of Autoencoders, Convolutional Neural Networks (CNNs), and Reinforcement Learning (RL), assessing their precision and computational efficiency across a range of maze complexities. In our exploration of CNNs, we conducted thorough testing across multiple iterations, involving an assortment of architectural models and variations in training paradigms. For RL, our experiments spanned focused on training a generalized DQN that can solve more than one maze. To establish a performance benchmark, these AI methods developed were evaluated against traditional graph search algorithms, notably Breadth-First Search (BFS), Depth-First Search (DFS), and A* Heuristic algorithms. The comparative results indicate that the Autoencoder approach outshines other evaluated AI techniques, demonstrating superior performance in maze-solving tasks.
Analysing the results of the benchmarking completed above, we firstly observe that with regards to the Solution Accuracy metric, amongst the AI driven methods, the Autoencoder performed the best, followed by the CNN and then the RL. We chose to exclude the algorithms from this metric as they offer guaranteed solutions to the type of maze puzzles within our project scope.