Simon Box

An automated signalized junction controller that learns strategies from a human expert

An automated signalized junction control system that can learn strategies from a human expert has been developed. This system applies machine learning techniques based on logistic regression and neural networks to affect a classification of state space using evidence data generated when a human expert controls a simulated junction. The state space is constructed from a series of bids from agents, which monitor regions of the road network. This builds on earlier work which has developed the High Bid auctioning agent system to control signalized junctions using localization probe data. For reference the performance of the machine learning signal control strategies are compared to that of High Bid and the MOVA system, which uses inductive loop detectors. Performance is evaluated using simulation experiments on two networks. One is an isolated T-junction and the other is a two junction network modelled on the High Road area of Southampton, UK. The experimental results indicate that machine learning junction control strategies trained by a human expert can outperform High Bid and MOVA both in terms of minimizing average delay and maximizing equitability; where the variance of the distribution over journey times is taken as a quantitative measure of equitability. Further experimental tests indicate that the machine learning control strategies are robust to variation in the positioning accuracy of localization probes and to the fraction of vehicles equipped with probes.

An automated signalized junction controller that learns strategies by temporal difference reinforcement learning

This paper shows how temporal difference learning can be used to build a signalized junction controller that will learn its own strategies through experience. Simulation tests detailed here show that the learned strategies can have high performance. This work builds upon previous work where a neural network based junction controller that can learn strategies from a human expert was developed (Box and Waterson, 2012). In the simulations presented, vehicles are assumed to be broadcasting their position over WiFi giving the junction controller rich information. The vehicles position data are pre-processed to describe a simplified state. The state-space is classified into regions associated with junction control decisions using a neural network. This classification is the strategy and is parametrized by the weights of the neural network. The weights can be learned either through supervised learning with a human trainer or reinforcement learning by temporal difference (TD). Tests on a model of an isolated T junction show an average delay of 14.12s and 14.36s respectively for the human trained and TD trained networks. Tests on a model of a pair of closely spaced junctions show 17.44s and 20.82s respectively. Both methods of training produced strategies that were approximately equivalent in their equitable treatment of vehicles, defined here as the variance over the journey time distributions.

Stochastic Six-Degree-of-Freedom Flight Simulator for Passively Controlled High-Power Rockets

This paper presents a method for simulating the flight of a passively controlled rocket in six degrees of freedom, and the descent under parachute in three degrees of freedom. Also presented is a method for modeling the uncertainty in both the rocket dynamics and the atmospheric conditions using stochastic parameters and the Monte Carlo method. Included within this, we present a method for quantifying the uncertainty in the atmospheric conditions using historical atmospheric data. The core simulation algorithm is a numerical integration of the rocket’s equations of motion using the Runge-Kutta-Fehlberg method. The position of the rocket’s center of mass is described using three dimensional Cartesian coordinates and the rocket’s orientation is described using quaternions. Input parameters to the simulator are made stochastic by adding Gaussian noise. In the case of atmospheric parameters, the variance of the noise is a function of altitude and noise at adjacent altitudes is correlated. The core simulation ...

Low cost infrastructure free form of indoor positioning

This paper is concerned with smartphone based inertial sensors for locating pedestrians. Two set of experiments were conducted over seven subjects. One set of the subjects walked along a straight path at a perceived normal and speedy pace, and the other set of subjects traversed a simple map consisting of two L shaped paths. In the experiments, a smartphone - HTC 2710e was handheld, with screen facing upwards was used to record the acceleration, magnetometer and gyroscope samples at 20Hz. The accelerometer samples are then analyzed to detect and count the footsteps, while gyroscope and magnetometer samples are used to estimate the heading. The results of our experiment show that the average error in foot step detection rate is less than 3% in normal walking and 5% in speedy walking using Fast Fourier Transform (FFT) based approach. While the most nearly correct estimated displacement is via Weiner approach with average percentage error of 5.74% in normal walking and 8.07% in speedy walking. The average percentage error in heading is less than 110 in turnings in both the L shaped paths.

An open-source, stochastic, six-degrees-of-freedom rocket flight simulator, with a probabilistic trajectory analysis approach

Predicting the flight-path of an unguided rocket can help overcome unnecessary risks. Avoiding residential areas or a car-park can improve the safety of launching a rocket significantly. Furthermore, an accurate landing site prediction facilitates recovery. This paper introduces a six-degrees-of-freedom flight simulator for large unguided model rockets that can fly to altitudes of up to 13 km and then return to earth by parachute. The open-source software package assists the user with the design of rockets, and its simulation core models both the rocket flight and the parachute descent in stochastic wind conditions. Furthermore, the uncertainty in the input variables propagates through the model via a Monte Carlo wrapper, simulating a range of possible flight conditions. The resulting trajectories are captured as a Gaussian process, which assists in the statistical assessment of the flight conditions in the face of uncertainties, such as changes in wind conditions, failure to deploy the parachute, and variations in thrust. This approach also facilitates concise presentation of such uncertainties via visualisation of trajectory ensembles.

Trajectory clustering, modelling, and selection with the focus on airspace protection

Take-off and landing are the periods of a flight where aircraft are most vulnerable to a ground based rocket attack by terrorists. While aircraft approach and depart from airports on pre-defined flight paths, there is a degree of uncertainty in the trajectory of each individual aircraft. Capturing and characterizing these deviations is important for accurate strategic planning for the defence of airports against terrorist attack. A methodology is demonstrated whereby approach and departure trajectories to a given airport are characterized statistically from historical data. It uses a two-step process of first clustering to extract the common trend, and then modelling uncertainty using Gaussian processes. Furthermore it is shown that this approach can be used to either select probabilistic regions of airspace where trajectories are likely and - if required - can automatically generate a set of representative trajectories, or select key trajectories that are both likely and critically vulnerable. An evaluation of the methodology is demonstrated on an example data-set collected by the ground radar at an airport. The evaluation indicates that 99.8% of the calculated footprint underestimates less than 5% when replacing the original trajectory data with a set of representative trajectories

Modeling the Dispersion of Aircraft Trajectories Using Gaussian Processes

This work investigates the application of Gaussian processes to capturing the probability distribution of a set of aircraft trajectories from historical measurement data. To achieve this, all data are assumed to be generated from a probabilistic model that takes the shape of a Gaussian process. The approach to Gaussian-process modeling used here is based on a linear expansion of trajectory data into a set of basis functions that may be parametrized by a multivariate Gaussian distribution. The parameters are learned through maximum-likelihood estimation. The resulting probabilistic model can be used for both modeling the dispersion of trajectories along the common flight path and for generating new samples that are similar to the historical data. The performance of this approach is evaluated using three trajectory data sets: toy trajectories generated from a Gaussian distribution, sounding-rocket trajectories that are generated by a stochastic rocket flight simulator, and aircraft trajectories on a given d...

A Gaussian process based decision support tool for air-traffic management

Technological developments in the last decade have shifted challenges in traffic flow management from obtaining and storing data, to analysing and presenting the enormous amount of available trajectory data in a comprehensible manner. This paper introduces a novel approach to visualising air-traffic, shifting the focus from displaying traffic density, towards directly visualising the flight corridors used by air-traffic. Such an approach is suitable for visualising air-traffic in three dimensions, which is particularly helpful in the vicinity of an airport where the air-traffic often changes level. Furthermore, the approach is data-driven, allowing the comparison of multiple trajectory datasets in order to identify changes in traffic corridors related to changing air-traffic and weather conditions. Finally, by using the probabilistic nature of the approach, it is possible to quantify the air-traffic complexity in terms of the traffic structure. The results presented in this paper show the approach applied to a trajectory dataset as measured by ground-radar near Denver airport (DEN).

Dataset for Modelling the dispersion of aircraft trajectories using Gaussian processes

Matlab code and trajectory data supporting: Eerland, Willem, Box, Simon and Sobester, Andras (2016) Modelling the dispersion of aircraft trajectories using Gaussian processes. Journal of Guidance Control and Dynamics.

Supplemental material for An open-source, stochastic, six-degrees-of-freedom rocket flight simulator, with a probabilistic trajectory analysis approach

Dataset supporting the Scitech 2017 conference paper.Predicting the flight-path of an unguided rocket can help overcome unnecessary risks. Avoiding residential areas or a car-park can improve the safety of launching a rocket significantly. Furthermore, an accurate landing site prediction facilitates recovery.This paper introduces a six-degrees-of-freedom flight simulator for small, solid-fuelled, unguided rockets. The open-source software package assists the user with the design of rockets, and its simulation core models both the rocket flight and the parachute descent in stochastic wind conditions.Furthermore, the uncertainty in the input variables propagates through the model via a Monte-Carlo wrapper, simulating a range of possible flight conditions.The resulting trajectories are captured as a Gaussian process, which assists in the statistical assessment of the flight conditions in the face of uncertainties, such as changes in wind conditions, failure to deploy the parachute, and variations in thrust.This approach also facilitates concise presentation of such uncertainties via visualisation of trajectory ensembles.