Learning for Agent-Based Systems

		 Slawomir Nowaczyk
		 University of Science and Technology (AGH)
		 Krakow, Poland

Classification is the most typical example of machine learning, but
definitely not the only one. When applying learning in an agent
setting, one needs to solve a different problem: each decision made by
the agent influences all the subsequent situations it will face. This
means the agent needs to not only take into account current
circumstances, but also all the future consequences of chosen action.

In my lecture I will talk about a number of methods of lifelong
learning of agents and different ways to approach the problem. By far
the most successful technique is reinforcement learning, although it
is not without flaws. Therefore other inductive learning approaches
are being developed, including various ideas concerning learning with
limited resources and combining learning with planning.