| dc.description.abstract | This study delves into the potential of quantum Deep Q-Learning (DQL) as an efficient approach to solving OpenAI Gym problems, setting it against the backdrop of classical DQL. Within the dynamic landscape of quantum computing, harnessing quantum properties such as entanglement and superposition in Quantum Neural Networks (QNNs). The project focuses on a selection of OpenAI Gym problems, Cart Pole, and Lunar Lander, and specific QNN architectures using Parameterized Quantum Circuits (PQCs) with re-uploading, combined with DQL algorithms. The methodology hinges on implementing and training classical and quantum DQL models, subsequently comparing their performance to highlight key distinctions. Emerging findings suggest that Quantum DQL (QDQL) can attain comparable performance to classical DQL, albeit with considerably fewer trainable parameters. However, it’s important to note that this performance may come with increased sensitivity to hyperparameters. A unique aspect of this study is the exploration of trainable entanglement as opposed to the static entanglement commonly addressed in the literature. Preliminary results indicate that a specific level of entanglement can benefit QNNs, although these findings remain inconclusive due to the short duration of training. Therefore, we propose a QDQL agent with trainable entanglement. In conclusion, this research underscores the potential of QDQL in reinforcement learning problems and the trade-offs involved, setting the stage for future investigations and optimizations of quantum computing approaches in machine learning. Future work could delve into a broader range of QNN architectures, alternative entanglement strategies, and diverse problem domains. | en_US |
