Russell Y. Webb

Reservoir Computing for Prediction of the Spatially-Variant Point Spread Function

A new method is presented which provides prediction of the spatially variant point spread function for the restoration of astronomical images, distorted by atmospheric turbulence when viewed using ground-based telescopes. Our approach uses reservoir computing to firstly learn the spatio-temporal evolution of aberrations caused by turbulence, and secondly, predicts the space-varying point spread function (PSF) for application of widely-used deconvolution algorithms, resulting in the restoration of astronomical images. In this article, a reservoir-based, recurrent neural network is used to predict modal aberrations that comprise the spatially variant PSF over a wide field-of-view using a time-series ensemble from multiple reference beacons.

The restoration of extended astronomical images using the spatially-variant point spread function

In this paper we address the problem of image restoration of extended astronomical objects, where the effects of inhomogeneous random media, such as atmospheric turbulence, result in distortion over the pupil plane. Distortions are usually expressed in terms of the point spread function (PSF) and can be used to restore an image with the application of a deconvolution algorithm. Since the PSF is generally spatially dependent, restoration of extended images within specific homogeneous (isoplanatic) regions is required. Inconsistencies at the borders of each region result in error boundaries between restored isoplanatic regions and the object. To minimise such distortion, we determine the optimal size of each isoplanatic region and then apply a linear filter to estimate the PSF between boundaries.

MULTI-LAYER CORRECTIVE CASCADE ARCHITECTURE FOR ON-LINE PREDICTIVE ECHO STATE NETWORKS

An architecture for on-line learning of time series prediction is presented which uses a series of echo state networks (ESNs). Each ESN learns to predict an error correction term for the previous ESN. This technique is demonstrated to improve prediction accuracy for on-line learning of the Mackey-Glass chaotic oscillator. The results are compared to other architectural configurations to show that the improved performance emerges from sequential ESN error correction. A new recurrent network structure is shown to be a useful simplification of the usual ESN reservoir.

Dynamic Artificial Neural Networks for Centroid Prediction in Astronomy

Motivation for this research is the real-time restoration of faint astronomical images through turbulence over a large field-of-view. A simulation platform was developed to predict the centroid of a science object, convolved through multiple perturbation fields, and projected on to an image plane. Centroid data were selected from various source and target locations and used to train an artificial neural network to estimate centroids over a spatial grid, defined on the image plane. The capability of the network to learn to predict centroids over new target locations was assessed using a priori centroid data corresponding to selected grid locations. Various distortion fields were used in training and simulating the network including data collected from observation runs at a local observatory. Results from this work provide the basis for extensions and application to modal tomography.

A simple metric for turn-taking in emergent communication

To facilitate further research in emergent turn-taking, we propose a metric for evaluating the extent to which agents take turns using a shared resource. Our measure reports a turn-taking value for a particular time and a particular timescale, or “resolution,” in a way that matches intuition. We describe how to evaluate the results of simulations where turn-taking may or may not be present and analyze the apparent turn-taking that could be observed between random independent agents. We illustrate the use of our turn-taking metric by reinterpreting previous work on turn-taking in emergent communication and by analyzing a recorded human conversation.

On the Probability of Improved Accuracy With Increased Sample Size

In this article, we consider the probability of improved accuracy in parameter estimation due to increases in sample size. For a range of scenarios we quantify this probability and develop a simple sample size formula to guarantee a certain level of improvement with a given probability. Our results lead to insights into sample size requirements and how they are related to the parameters governing the sampling process.

Low temperature differential stirling engine based power generation

Research into various forms of sustainable power generation is required to address the rising problems associated with fossil fuel limitations. A new research programme aims to show that electric power generation using low-grade heat and Stirling engine technology can be made commercially viable. Using the Modified Beale Number, and preliminary software modelling, an initial prototype has been designed. This prototype will output approximately 1 kW of electric power with a temperature differential of around 30 K. The fundamental considerations for this prototype are the effects of engine parameters such as compression ratio, pressure, regenerator matrix composition, displacer/power piston phasing, and displacer velocity profile. The result here being a novel gamma-type engine configuration that complements a high pressure system. The prototype also has an electrically actuated displacer piston so that the phasing and velocity profile are finely variable, and a modular regenerator matrix.

Rewards for pairs of Q-learning agents conducive to turn-taking in medium-access games

We describe a class of stateful games, which we call ‘medium-access games’, as a model for human and machine communication and demonstrate how to use the Nash equilibria of those games as played by pairs of agents with stationary policies to predict turn-taking behaviour in Q-learning agents based on the agents’ reward function. We identify which fixed policies exhibit turn-taking behaviour in medium-access games and show how to compute the Nash equilibria of such games by using Markov chain methods to calculate the agents’ expected rewards for different stationary policies. We present simulation results for an extensive range of reward functions for pairs of Q-learners playing medium-access games and we use our analysis for stationary agents to develop predictors for the emergence of turn-taking. We explain how to use our predictors to design reward functions for pairs of Q-learning agents that are conducive (or prohibitive) to the emergence of turn-taking in medium-access games. We focus on designing multi-agent reinforcement learning systems that deliberately produce coordinated turn-taking but we also intend our results to be useful for analysing emergent turn-taking behaviour. Based on our turn-taking related results, we suggest ways to use our methodology to designs rewards for quantifiable behaviours besides turn-taking.

Low enthalpy heat stirling engine based electric power generation: A research design

Power generation from renewable and sustainable sources is an important field of technology development owing to the increasing costs (both economically and environmentally) of fired generation. Most research associated with alternative forms of power generation have concentrated on relatively high efficiency systems, as this makes commercial realisation more readily achievable. A new research programme is taking a slightly different approach, where an inherently low efficiency power generation system based on low temperature differential Stirling engine technology is being considered. Here we describe the design of a low temperature differential Stirling engine for research into determining the commercial viability of utilising low enthalpy (or low grade) heat for electric power generation. While the design does not satisfy low cost economic requirements, through its ability to change important operational parameters such as compression ratio, piston phasing, and working gas movement dynamics, it is capable of identifying important features needed to optimise engine performance. The design utilises the Modified Beale Number and some preliminary software modelling, to satisfy the specifications of having a 500W output with a temperature differential of only 40 K.

Time Series Prediction with Evolved, Composite Echo State Networks

A framework for predictive, on-line, learning networks composed of multiple echo state networks is presented. These composite networks permit learning predictions based on complex combinations of sub-predictions and error terms. The configuration space is explored with a genetic algorithm and better performance is achieved than with hand coded solutions.