CausalXRL

Artificial intelligences (AI) are now capable of solving complex tasks at a level that until recently was reserved for natural intelligences. These tasks include the masterful playing of chess and Go, but also autonomous driving or the diagnosis of medical diseases. However, AI systems require a great deal of practice to master these tasks. For instance, it may require several million attempts until an AI system has learned which moves in chess lead to desired future playing positions. Such extensive training is only possible if the AI system`s environment can be simulated on a computer, such as in chess. If the AI system must operate in a real, physical environment, such as in autonomous driving or in assisting physicians in the diagnosis and treatment of diseases, the AI system lacks a fundamental human ability to correctly assess the consequences of its own actions from a few trials - the ability to infer causal relationships in the environment from pure observation. Closely linked to this skill is the ability to explain and justify it`s actions. The goal of the CausalXRL project is to equip AI systems with these abilities. Thus, the CausalXRL project paves the way to use AI systems in scenarios where they need to learn efficiently and justify their decision to human supervisors in an intuitively tangible way, such as in the support of physicians in the diagnosis and treatment of diseases.

Artificial intelligences (AI) are now capable of solving complex tasks at a level that was previously reserved for natural intelligences. These tasks include masterfully playing chess and Go, as well as autonomous driving or diagnosing medical conditions. However, to successfully master these tasks, AI systems require a great deal of practice. It often takes several million attempts for an AI system to independently learn which moves in chess lead to desired future game positions. Such extensive training is only possible when the environment of the AI system can be simulated on a computer, as is the case with chess. When the AI system must operate in a real, physical environment, such as in autonomous driving or in assisting doctors in the diagnosis and treatment of diseases, it lacks a fundamental human ability to correctly assess the consequences of its own actions based on a few trials - the ability to infer causal relationships in the environment from mere observation. Closely related to this is the ability to explain its own actions and justify them in such a way that they become understandable to other natural or artificial agents. The CausalXRL (Causal Explanations in Reinforcement Learning) project has made fundamental contributions to equipping AIs with these very characteristics. New algorithms have been developed in the project that reduce the complexity of (natural or artificial) neural computations of intelligent agents to a few, meaningful concepts in such a way that these still provide a causally sufficient explanation for the agents' behavior. Through this simplification, the CausalXRL project has helped lay the theoretical and algorithmic foundations for AIs and humans to understand each other in a way that allows them to cooperatively tackle complex challenges."