Monday 4/26/93 ;  3:00 ; WeH 7220

Sven Koenig -- Complexity Analysis of Reinforcement Learning



In this talk, I will give a short overview of results about the
complexity of real-time reinforcement learning in deterministic
domains and how it can be decreased. 

In particular, I will discuss the influence of ...

- the way the domain is represented
- the initial Q-values (or U-values)
- domain properties, for example:
     one can undo every action immediately ("no one-way streets")

and will then focus on the influence of the reactivity of the
algorithm, i.e. how much planning is done between action executions.

I will also discuss what deterministic and probabilistic domains have
in common and how they differ, thereby outlining future research.

This is NOT a theory talk, although it is not empirical stuff either.
It is meant to provide empirical researchers with guidelines on how to
represent their domains, which domain properties to watch out for etc.