Design of Autonomous Agent that Adapts and Grows in Complexity

Background:

Most robots are designed to increase their learning with regard to single goals but are not capable of developing new goals. For example, a machine designed to explore terrain may continue to explore and become very effective at discovering previously unknown areas but will never write a sonata. These robots that are built on maximizing one reward function are likely to have problems performing in changing or highly complex environments.

Technology Overview:

Researchers at the University at Albany have developed an algorithm that enables a robot that has been programmed to carry out simple, concrete, and short-term goals to develop into one that has a complex motivational structure and can pursue abstract and long-term goals. In addition to enabling a machine to learn to reach its goals, the algorithm allows it to minimize conflict between goals and adapt to unforeseen circumstances and dynamic situations—all without the need for a supervisor. No similar algorithms appear to exist, despite the recognition of the significant limitations of existing algorithms and a strong need for devices that can self-develop.

Advantages:

•    Advances the capabilities of simple robots to respond to complex situations
•    Algorithm is highly adaptive, enabling robots learn from a wide variety of problems

Applications:

 List of potential uses or markets.

Intellectual Property Summary:

•    Know-how
•    Copyright

Licensing Status:

This technology is available for licensing.