Henrik Berg

Automatic design of pulse coupled neurons for image segmentation

Automatic Design of Algorithms through Evolution (ADATE) is a program synthesis system that creates recursive programs in a functional language with automatic invention of recursive help functions and self-adaptive optimization of numerical values. We implement a neuron in a pulse coupled neural network (PCNN) as a recursive function in the ADATE language and then use ADATE to automatically evolve better PCNN neurons for image segmentation. Our technique is generally applicable for automatic improvement of most image processing algorithms and neural computing methods. It may be used either to generally improve a given implementation or to tailor that implementation to a specific problem, which with respect to image segmentation for example can be road following for autonomous vehicles or infrared image segmentation for heat seeking missiles that are to distinguish the heat source of the target from flares.

Automated synthesis of control algorithms from first principles

A variety of machine learning techniques have been employed to automatically create control algorithms for autonomous vehicles. Much research has focused on various ¿black box¿ approaches, in which the synthesized or learned control algorithms perform well when tested, but are difficult or impossible to analyze and understand. This paper presents the use of the ADATE system to evolve a control algorithm based on a racing car simulator. The system evolved compact and analyzable yet sophisticated control algorithms capable of driving millions of randomly generated tracks at high speeds without ever driving off the road. The approach presented is likely to be applicable to most automatic control problems, given a set of training examples and a suitable software simulator.

Viewing the problem from different angles: a new diversity measure based on angular distances

It is commonly believed that diversity is crucial for an evolutionary system to succeed, especially when the problem to be solved contains local optima from which the population cannot easily escape. There exist numerous methods to measure population diversity, but none of these have been shown to be consistently useful. In this paper, a new diversity measure is introduced, and it is shown that high diversity according to this new measure generally leads to a more successful overall evolution in most of the cases considered.

A comparison of different machine learning algorithms for automatic classification of sonar targets

A well-known problem with modern anti-submarine warfare sonars with narrow beamwidths and wide frequency bandwidths, is the frequent occurence of false alarms, particularly in littoral environments. This increases the workload of sonar operators and also reduces the usefulness of automatic systems such as autonomous underwater vehicles, since their limited communication abilities hinder them from sharing large amounts of contacts. In this paper, four traditional machine learning algorithms are tested on sonar data with a high amount of false alarms together with synthetic submarine echoes. It is shown that some of the algorithms can outperform simple signal to noise ratio (SNR) thresholding by a significant amount, but that the performance is highly dependent on the parameter values chosen for each algorithm. These parameters are therefore investigated in order to determine their relative significance.

Synthesis of Control Algorithms for Autonomous Vehicles through Automatic Programming

In this paper, an automatic approach to designing a control system is presented. An algorithm for the control of a simulated race car is evolved and shown to be capable of driving millions of randomly generated tracks at high speeds without ever driving off the road. We also show that our automatically generated non-linear control algorithm significantly outperforms a linear model. Although the application considered is that of controlling a racing car, our methodology is likely to be generally applicable to most automatic control problems given suitable software simulators for the systems in question.

The classification performance of signal-to-noise ratio and kinematic features in varying environments

The use of high resolution, broadband sonars during anti-submarine warfare operations often result in a sonar image cluttered with false alarms. This may prove challenging for a sonar operator when it comes to correctly distinguishing sonar contacts from false alarms. These false alarms may originate from different sources such as seamounts, underwater ridges, topographical features in general or man-made objects, such as pipelines. The unmanagable amounts of false alarms are today handled most commonly by raising the detection threshold on the signal-to-noise ratio. This would, however, also lower the probability of detecting a valid target, such as a submarine. At track level, false alarms from topographical features, such as seamounts, should be very different from those resulting from pipelines, hence, leading to differences in their kinematic properties. Machine learning should then be able to exploit these different sources of information in order to separate actual submarine detections from false alarms. However, it is well known that the application of such methods on a dataset from one environment is not necessarily valid for a different environment. Here we apply the above mentioned method on two fundamentally different datasets. The first dataset, Clutter Experiment 2 (CEX02), was collected close to the coast where the false alarms mainly consist of returns from a rocky, upslope bottom. The second dataset, Localization Experiment 2 (LOC02), is from the middle of the Norwegian trench, an area characterised by a very flat bottom and far fewer false alarms. Both datasets were collected in 2002 during the New Array Technology III (NAT III) programme, which was a combined effort between the Norwegian Defence Research Establishment (FFI), Thales Underwater Systems, TNO, the Royal Norwegian Navy, the Royal Dutch Navy, and the French Navy. The trial was carried out in the Norwegian Trench where a low-frequency source and an array receiver were towed by FFIs research vessel, HU Sverdrup II. The data are processed up to track level before the kinematic properties and signal-to-noise ratio of both false tracks and target tracks are extracted. Due to the confidential nature of submarine tracks these are removed from the analysis and pipeline tracks are considered as target tracks. An important find was that automatic classifiers trained through machine learning on data from one environment does not necessarily work in a different environment. Furthermore, we observe that classifiers tested on data from the same location as the training dataset exhibit signs of overtraining. Finally, the best performance in all instances is found when the classifiers are determined from a combined dataset of both environments before applied to the different environments separately.

Fitness Sharing Based on Angular Distances

It is commonly believed that diversity is crucial for an evolutionary system to succeed, especially when the problem to be solved contains local optima from which the population cannot easily escape. There exist numerous methods to maintain the diversity of an evolving population, but it is not always clear what kind of diversity is helpful in a given situation. In this paper we show that striving to maintain high angular distances between the fitness vectors of the individuals in a population leads to better results in most cases considered. Without increased computational costs, our angular sharing scheme enables the evolutionary system in most cases to find better solutions than other sharing schemes investigated.

Improving the scalability of automatic programming

When automatically constructing large programs using program transformations, the number of possible transformations grows very fast. In this paper, we introduce and test a new way of combining several program transformations into one transformation, inspired by the combinatorial concept of Covering Arrays (CA). We have equipped the ADATE automatic programming system with this new CA transformation and conducted a series of 18 experiments which show that the CA transformation is a highly useful supplement to the existing ADATE transformations.