Year: 2024
Type: dissertation
Abstract: The board game Diplomacy presents a rich and challenging proving ground for negotiating agents, as players may choose to further their goals through honest or deceitful communication. This thesis examines whether an agent can gain an advantage by sending intentionally deceptive messages. To this end, a deceptive agent is developed based on Cicero, a Diplomacy agent created by Meta AI. A theoretical analysis of Cicero reveals how its approximation of an optimal policy can be improved, motivating specific design principles for a deceptive agent. Following these principles, a conditional language model is repurposed to generate intentionally deceptive messages, with effects predicted by extrapolating from truthful dialogues between human players. Selected scenarios from a series of simulated games demonstrate that the agent can produce convincing and strategically beneficial deceptions, validating its design principles.
Author: Janson, Axel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: strategic communication, negotiating agents
Open Access status: green
OpenAlex ID: W7053002460
