ABIAS

Not all data is equally useful. A major challenge in training artificially intelligent systems that learn from interactions with their environments (agents), is to acquire the most useful data points. For example, where should a robot look in order to pick up a cup? Active inference is a framework for designing agents that balance information-seeking and goal-seeking behaviour. This project will dive into the information-theoretic basis of this framework. We will attempt to answer questions such as:

• How close to optimal is data acquisition based on maximizing information gain / minimizing expected free energy?
• Under what conditions is the active inference agent a consistent parameter estimator?
• Under what conditions is the active inference agent a stable controller?

This position is supported by the Sectorplan Techniek of the Dutch Ministry of Education, Culture and Science.