Transfer of Knowledge in Reinforcement Learning through Policy-based Proto-Value Functions

Transfer learning is a process that occurs when learning in one context affects the performances of learning in another contexts. On the other hand,
reinforcement learning is a very general learning paradigm, which potentially
can be used to solve any learning and control problem. Nevertheless it suffers from high complexity problems. Transferring knowledge across dierent
tasks can be important, since it makes possible to re-use existing experience
in order to reduce the complexity both in terms of required experience and
time.
There have already been several proposals of transfer learning techniques
inside reinforcement learning. However, these consist of separate eorts,
which on a rst sight appear to have few points of contacts.
In this thesis we give two main contributions. The rst contribution is to
provide a sketch for a unied framework of transfer learning in reinforcement learning, identifying three dierent sub-objectives for transfer. Among
these three objectives, we propose solution strategies for what concerns one
of them: representation transfer, i.e. learning and transfer of representation across tasks. Our approach is based on proto-value functions, a set of
techniques that enables learning of the domain representation via learning
of basis functions for function approximation.