Featured Senior Project Expose: Conversation with Sam Arnesen Class of 2020
In the upcoming spring semester, senior computer science major Sam Arnesen is set to focus on his thesis, which explores the development of artificial intelligence software that can solve text-based computer games. Arnesen's unique project is a departure from his previous work in computer vision and is inspired by his adviser's work on this topic.
The challenges associated with text-based games are numerous, including the partially observed nature and the difficulty in parsing text and translating it into meaningful actions. Arnesen's thesis aims to address these challenges by focusing on ways to have more transfer of learning between different kinds of text-based games.
Arnesen's spring semester goal for his thesis is not to leave it to the last moment, and he has identified the need to devote more time to his work. He advises future seniors to start early on their independent work, be honest about their workload, and not to overcommit during the fall semester.
The key approaches in Arnesen's thesis work include successor features for knowledge transfer, generalized policy improvement algorithms, natural language processing (NLP) techniques, and transfer learning approaches. By combining these methods, Arnesen's AI agent will be able to adapt to different game contexts more effectively.
Arnesen's approach involves building an agent that parses walk-throughs of games. This agent will use NLP tools like SpaCy to understand player inputs and game narratives more effectively, supporting modular interpretation of commands across different game contexts.
While substantial work has focused on visual and strategic games, text-based games require specialized attention to language grounding and understanding, which is key for successful transfer. Arnesen's work aims to fill this gap by improving the AI’s ability to transfer learning between different types of text-based games.
Arnesen's thesis work will not involve sharing it on Reddit or being part of the Seasonal Series: An Interview with Eric Ahn. However, he expects his work to receive more attention compared to the first semester. Arnesen aims to finish his thesis earlier in the spring semester compared to the first semester.
References:
[1] Sutton, R. S., & Barto, A. G. (1998). Reinforcement learning: An introduction. MIT Press.
[2] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. R., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., & Hassibi, B. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529-533.
[3] Yudkowsky, E. (2010). Artificial General Intelligence as a Cognitive System. In C. H. W. L. P. H. M. J. A. M. (Eds.), Artificial General Intelligence: A Roadmap (pp. 1-13). Springer.
Sam Arnesen, in his spring semester, intends to delve deeper into his senior thesis, which revolves around developing an AI software that can tackle text-based computer games. He plans to focus on enhancing transfer of learning between various text-based games, employing strategies such as successor features, generalized policy improvement algorithms, NLP techniques, and transfer learning approaches, along with leveraging tools like SpaCy for natural language processing. To ensure a successful outcome, Arnesen advises future seniors to begin their independent work early, manage their workload honestly, and resist overcommitting during the fall semester.
