Objective
The topic of this research is robust machine learning in the face of many conflicting tasks and objectives the agent is responsible for learning. As learning algorithms become a more important component of future systems, the flexibility required for these algorithms to address the ever-increasing range and complexity of tasks we ask of them will require advancements in machine
learning techniques. It is the view of the researcher that one very important aspect of this advancement will be the development of techniques that can handle an order of magnitude or more greater concurrent tasks without requiring undue effort on the part of human programmers or unreasonable computation time or computational burden. This work will develop more flexible learning mechanisms that can allow for the degree of autonomy that will be necessary in the next
generation of learning systems. In particular, improvements are necessary in life-long learning behavior of agents that are intended to address a wider range of problems than most current systems are capable of, and such systems need to be able to concurrently manage many more goals and tasks than they currently do, and do so with less direct human intervention. In particular, this research focuses on the problems inherent in agents which must maintain the goal of
learning very many heterogeneous tasks simultaneously. These problems primarily lie in scaling the learning algorithms, in more intelligent management of action selection to ensure that the agent is able to balance the conflicting goals without ignoring any of the tasks. The research proposes to address the problems by adapting current research from the field of multi-objective optimization to
help develop new scalable reinforcement learning algorithms. In addition, new algorithms will be developed to perform action selection in the presence of many heterogeneous goals.
Fields of science
Call for proposal
FP7-PEOPLE-2011-CIG
See other projects for this call
Coordinator
101 REYKJAVIK
Iceland