Gaming, Goats & General Intelligence with Frederic Besse
Google DeepMind: The Podcast
Developing Intelligent Game Agents through Imitation Learning
Building intelligent game agents involves partnerships with game developers to utilize specific 3D games, allowing for custom objectives defined through language. Agents are trained using imitation learning by collecting data from human players. This approach relies on the agents mimicking human behavior without traditional reward systems, generating probabilistic actions based on observed player choices. The use of games like Valheim and Goat Simulator showcases how agents learn to perform tasks such as gathering resources or causing chaos within a flexible environment. The unpredictability in agents' actions can lead to both entertaining and unexpected outcomes, as demonstrated in Goat Simulator, where agents may adopt humorous, unintended behaviors during playtesting.
Remember Everything You Learn from Podcasts
