Abstract:

Agents, defined as programs or robots that interact with their environment, are becoming increasingly common. However, the current generation of agents are often unable to robustly interact with each other, or with humans, severely limiting the number of tasks that they can accomplish. Furthermore, these agents typically are unable to adapt to their environment, a critical skill when programmers do not have full knowledge of agents' future environments or when the agents' environment may change over time. This talk will discuss recent work in combining autonomous learning of sequential decision making tasks with transfer learning, a general approach to sharing knowledge between agents with different capabilities, resulting in significant improvements to learning speeds and abilities.