|
  |
|
|
Krasnow Institute > Monday Seminars > Abstracts Towards Developing an Entity That Learns High-level Cognitive Skills in a Lifelike Way Richard E. Neapolitan
Since 1970 the Artificial Intelligence Community has largely focused on developing domain-specific knowledge-based systems. Arguably their greatest success has been the development of the belief network and methods for doing inference in belief networks. Most recently, they have developed algorithms that learn belief networks from statistical data. Although this research has resulted in powerful tools for creating systems that do medical diagnosis, language understanding, planning, etc., it has had little impact on the development of artificial intelligence. That is, if we require an intelligent agent to be capable of rapidly interpreting its environment, we have made little progress in the development of such an agent. The algorithms that learn belief networks are based on normative principles and do not learn incrementally. Once a knowledge-based system is developed, it can perform powerful intellectual tasks, but it cannot further develop by interacting with an environment. Recently, some biologists and philosophers have taken a different approach to artificial intelligence, which could be called the artificial life' approach. These researchers say that the problem with the Artificial Intelligence Community is that they focus on creating a disembodied intelligence. Alternatively, artificial-life researchers search for algorithms that describe the process by which a life-form learns' while interacting with its environment. One such algorithm is used in a robot that learns to bring light to the center of its retinal image due to an innate value for light. The idea is that innate and learned values are the basis by which a life-form develops while interacting with its environment. In general, the artificial-life approach has been applied to developing entities that learn crude skills such as seeking light and not bumping into walls. Some in the Artificial Intelligence Community maintain that humans structure causal knowledge in a belief network, and that they perform inference using that knowledge in the same way as a well-known belief network algorithm. This is in agreement with some philosophers and psychologists, who maintain that the fundamental unit of knowledge is the cause-effect' relation. However, the Artificial Intelligence Community has not addressed how the human learns the belief network in the first place. Again, their learning algorithms are based on normative principles and are not incremental. This talk describes the belief network, summarizes the arguments that it is a model of the way life-forms represent knowledge, suggests ways to test this hypothesis, describes some of the artificial-life algorithms for learning in a lifelike way, and addresses the question as to whether we can use these or similar algorithms to learn a belief network. It concludes by suggesting that we combine the methods of the Artificial Life and Artificial Intelligence Communities in an effort to develop an entity that can learn high-level cognitive skills by interacting with its environment.
The Krasnow Institute for Advanced Study |
|
