AI for playing Codenames board game using different language models
Creative computer systems grapple with challenging tasks that exist within effectively endless combinatorial spaces. Further complicating these already difficult tasks is the fact that the goal of high-quality creative output is itself nebulous. A creative domain with concrete goals would therefore be a fruitful domain for studying computer creativity. We propose that competitive language games are just such a domain---they require creativity but also feature concrete win and loss states. We present an analysis of creative agents that play one such game: Codenames, a 2016 board game of communicating hidden information via single-word hints. Our model-agnostic framework allows us to compare agents that utilize different language models. We present our findings and discuss how future computational creativity research can continue to explore competitive language games.
The associated paper, Competitive Language Games as Creative Tasks with Well-Defined Goals, was accepted for publication at ICCC 2022