Richard J. Cant

A framework for the evaluation of volume rendering techniques on a task specific basis using neural networks

Abstract This paper tests the use of neural networks to evaluate the effectiveness of volume rendering techniques. The advantage of using neural networks is that they are able to give a task-specific evaluation of an algorithm. In a practical application a specific training and testing regime would be constructed. This could then be used to decide on the appropriate algorithm for the application. In order that this methodology itself can be evaluated, a number of volume rendering algorithms are implemented and several different “tasks” are defined. For each task/algorithm combination a neural network is trained and tested. These results provide an evaluation of the algorithm for that task. Because the purpose of this paper is to investigate the neural network evaluation methodology itself, not the algorithms, the techniques chosen include some simplistic ones that have well understood deficiencies, to provide an objective test. It was found that the neural net methodology clearly detected the deficiencies, and so could be considered an effective framework for calibrated testing of problem-specific algorithms.

Methods for Automated Object Placement in Virtual Scenes

This paper introduces a variety of heuristic methods of object placement that can be used within a general scheme for the automatic creation of virtual rooms. The methods allow a room to be populated with objects in a randomised but controllable way, where the choice of method is determined by the object being placed and the target location. Certain combinations of object and location (for example books on a bookshelf) allow very simple methods to be used whilst most general situations are addressed using Stochastic Diffusion Search. The system is integrated with a Physics Engine, to guarantee physically stable configurations. These methods allow naturalistic placement to occur, for example paper within desk-tidy.

A hybrid Artificial Intelligence approach with application to games

We describe a hybrid Artificial Intelligence (AI) approach combining soft AI techniques (neural networks) and hard AI methods (alpha-beta game tree search), in an attempt to approximate human play more accurately, in particular with reference to the game of Go. The program is tested and analysed by play against another Go playing program and it is shown that the use of hard AI enhances the performance of the soft AI system and vice-versa.

On Practical Automated Engineering Design

In engineering, it is usually necessary to design systems as cheap as possible whilst ensuring that certain constraints are satisfied. Computational optimization methods can help to find optimal designs automatically. However, it is demonstrated in this work that an optimal design is often not robust against variations caused by the manufacturing process, which would result in unsatisfactory product quality. In order to avoid this, a meta-method is used in here, which can guide arbitrary optimization algorithms towards more robust solutions. This was demonstrated on a standard benchmark problem, the pressure vessel design problem, for which a robust design was found using the proposed method together with self-adaptive stepsize search, an optimization algorithm with only one control parameter to tune. The drop-out rate of a simulated manufacturing process was reduced by 30 % whilst maintaining near-minimal production costs, demonstrating the potential of the proposed method.

Prediction of Jet Engine Parameters for Control Design Using Genetic Programming

The simulation of a jet engine behavior is widely used in many different aspects of the engine development and maintenance. Achieving high quality jet engine control systems requires the iterative use of these simulations to virtually test the performance of the engine avoiding any possible damage on the real engine. Jet engine simulations involve the use of mathematical models which are complex and may not always be available. This paper introduces an approach based on Genetic Programming (GP) to model different parameters of a small engine for control design such as the Exhaust Gas Temperature (EGT). The GP approach has no knowledge of the characteristics of the engine. Instead, the model is found by the evolution of models based on past measurements of parameters such as the pump voltage. Once the model is obtained, it is used to predict the behaviour of the jet engine one step ahead. The proposed approach is successfully applied for the simulation of a Behotec j66 jet engine and the results are presented.

Creating Depth of Field Effects without Multiple Samples

We describe the depth of field A buffer, a new method for generating depth of field effects in synthesised images. The method avoids the use of multiple discrete samples but still has the capability to produce the effect of a sharp object viewed through a de-focussed one. Initial rendering can be done via a system producing pixel contributions together with depth information as required by conventional Z buffering. We present sample images illustrating the advantages of the technique compared to multi-sampling. We discuss the options for hardware-accelerated implementation and possible applications.

Efficient anti-aliased bump mapping

Abstract The naive method of anti-aliasing bump mapping, using MIP maps, does not generate theoretically correct results due to the non-linearity of the bump mapping process. In some cases results can be improved by modifying the resultant average beam to take account of the angular spread of the vectors being averaged. More generally it is necessary to represent each texel in a MIP map by multiple vectors. The resulting structure can be thought of as a multiple surface. A previous attempt to do this encountered problems with excessive time being required to generate the MIP maps. We present a new method of multi-vector MIP map generation in which the vectors are obtained by a competitive learning process. This method is orders of magnitude faster. The multi-vector technique is also extended to environment map based bump mapping where it is shown to be visually superior to earlier techniques. The method can be readily incorporated into current and future graphics cards with low cost in memory and processing.

Characterisation of Large Changes in Wind Power for the Day-Ahead Market Using a Fuzzy Logic Approach

Wind power has become one of the renewable resources with a major growth in the electricity market. However, due to its inherent variability, forecasting techniques are necessary for the optimum scheduling of the electric grid, specially during ramp events. These large changes in wind power may not be captured by wind power point forecasts even with very high resolution numerical weather prediction models. In this paper, a fuzzy approach for wind power ramp characterisation is presented. The main benefit of this technique is that it avoids the binary definition of ramp event, allowing to identify changes in power output that can potentially turn into ramp events when the total percentage of change to be considered a ramp event is not met. To study the application of this technique, wind power forecasts were obtained and their corresponding error estimated using genetic programming and quantile regression forests. The error distributions were incorporated into the characterisation process, which according to the results, improve significantly the ramp capture. Results are presented using colour maps, which provide a useful way to interpret the characteristics of the ramp events.

Fourier Texture Filtering

Many of the best texture filtering techniques are unsuitable for modern shader based graphics hardware, either because of performance considerations or simply because they do not fit in with the way the hardware operates. Those that are compatible with modern hardware often suffer other shortcomings. This paper presents a new texture filtering technique; which aims to supply similar visual quality to the top techniques, with smaller source textures, in real-time and on current consumer level hardware. This new technique utilises Fourier space filtering to minimise aliasing in MIP maps, while adding back detail by re-introducing the more dominant frequencies when the texture is rendered.

Entropy Measurement within Graphical Scenes

This paper proposes a method by which virtual scenes may be quantified in terms of their level of disorder. We have derived an entropy measure that is based on the physical and semantic nearness and relative orientations of the objects in the scene. Entropy measurement allows the detection and classification of player activity in games with realistic physics and large object counts. It also permits the performance of automated placement algorithms to be measured and optimised. Measurement is sufficiently quick to be able to occur during a game, enabling it to be used to aid the game to react to user behaviour in real time.