Erlangen 2018 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
Q: Fachverband Quantenoptik und Photonik
Q 16: Quantum Information and Simulation
Q 16.4: Talk
Monday, March 5, 2018, 14:45–15:00, K 1.020
Projective simulation applied to non-Markovian problems — •Lea M. Trenkwalder1, Vedran Dunjko2, and Hans J. Briegel1 — 1University of Innsbruck — 2MPI for Quantum Optics
The idea of machines acquiring complex behaviour can be studied in terms of learning agents. In reinforcement learning models, an agent learns through interaction with an environment, as it receives rewards and information about the environment in terms of percepts. An agent is confronted with a Markovian task environment if a given percept contains all the information needed to determine the probability distribution over the subsequent environmental states. Projective simulation (PS) is a novel learning model, which has been used to solve a variety of Markovian reinforcement learning tasks. The projective simulation model is a physics-inspired approach where the internal deliberation process of the agent can be described by a random walk through its episodic memory. Moreover, this random walk possesses a quantum analogue, providing the PS framework with a natural route to quantisation. For a variant of PS called rPS, it was proven that quantum effects can be exploited to achieve a quadratic speed-up in its active learning time. Recently, it was shown that complex Markovian task environments such as the design of certain quantum experiments, can be tackled using PS. In the present work, we generalise projective simulation to solve a set of non-Markovian problems, in which the perceptual input does not enclose all the information needed to determine the development of the environment. The approach allows the projective simulation model to be applied to a wider range of task environments.