This will be a very informal discussion/overview of recent work I've been doing using techniques directly from the Sutton/Barto book.
Consider an active agent that is learning about and behaving in an environment which it can influence by its own actions. Different actions will lead to different future observations, and intuitively, some action strategies should make it easier than others for the agent to construct a model that generalizes well to future observations. We take a novel approach to behavioral learning, based on this intuition. Formally, our approach is based on Shannon's mathematical theory of communication. We develop a theoretical framework that extends rate-distortion theory and the related Information Bottleneck method to a feed-back situation. This framework also allows us to take a fresh look at exploration in reinforcement learning (RL) and, in particular, at curiosity driven RL. One central result is that the optimal action strategies are characterized by a balance between exploration and control. In the case of reinforcement learning, when a goal is specified and optimized, exploration is balanced with exploitation. This balance emerges naturally from the information theoretically motivated principle without the need of extra heuristics.
Small-world networks have important applications in science and technology. Recent renormalization group results predict non self averaging behaviour at criticality for relevant disorder. However, we find strong self averaging (SA) behaviour in the critical region of a quenched Ising model on an ensemble of small-world networks, despite the relevance of random bonds at the pure critical point.
The Asymmetric exclusion process (ASEP) is one of the paradigms of non-equilibrium statistical physics where the boundary conditions play a non-trivial role in determining the stable phase diagram of open systems, as
opposed to equilibrium systems, where conditions in the bulk are paramount. We show that the introduction of disorder in the recent Fermionic generalisation of ASEP markedly changes the phase diagram by introducing an extra phase and changing the the multicritical point.
Thermal ratchets and Brownian motors are known to have important applications in science, engineering and even in game theory. However, Brownian motors exhibit sub-percentage efficiency. We present a prescription to achieve enhanced thermodynamic efficiency of energy transduction in Brownian motors and show that an efficiency close to the ideal limit can be obtained in the deterministic regime.
Noise-induced transport is inevitably beset with diffusion. This can lead to a loss of the ratcheting effect in systems with finite spatial extensions. We show that the suppression of backward currents can lead to
an extremely reliable transport (quantified by the Peclet number) in rocked ratchets. The values of Peclet number so attained are the highest in theoretical studies on Brownian motors till date.
John Mahoney will be giving a practice talk for his upcoming oral exam on Quantum Finite State Machines.
Time: 11:00am - Noon+ Place: KeckCAVES, Basement of Physics/Geology Building
Time: 11:00am - Noon+ Place: KeckCAVES, Basement of Physics/Geology Building
Detection, Discovery, and Analysis of Structure in High-Dimensional, Time-Delay Dynamical Systems: A Survey of Recent Results
Deducing, inferring and understanding the dynamics and geometric structures in high-dimensional, time-dependent systems has been, and will continue to be, a problem of considerable interest and importance to engineers, natural scientists, and mathematicians. The primary goal of this survey is to present various means of deducing geometrical structure and organization in high-dimensional, time-delay, chaotic dynamical systems that cannot be reduced to low-dimensional analogs. In this circumstance, the rescaling of time in conjunction with adding dimensionality can have an effect on both the fundamental properties of the dynamical system and on many common diagnostics used to study such systems. Because of the interaction between rescaling time and adding dimensions, both the effects of time/dimension on the geometric structure of the system (or ensemble of systems) as indicated by the diagnostics, as well as the effects of time rescaling on the diagnostics, will be discussed. General conclusions include: an invariance of the entropy to time-rescaling; persistence of chaotic dynamics as the dimension is increased; dimension calculations may yield deceiving results; and a tendency towards a continuous-LCE-spectrum dynamic as time-delay coordinates are added. Finally, preliminary results regarding the rescaling of LCEs via homogeneous functions is presented.
Please note the NEW TIME AND DAY of the group meeting: Tuesdays, 2:00 - 3:00pm, MSB 1106