Automatic Text Simplification (ATS) is one of the major Natural Language Processing (NLP) tasks, which aims to help people understand text that is above their reading abilities and comprehension. ATS models reconstruct the text into a simpler format by deletion, substitution, addition or splitting, while preserving the original meaning and maintaining correct grammar. Simplified sentences are usually evaluated by human experts based on three main factors: simplicity, adequacy and fluency or by calculating automatic evaluation metrics. In this paper, we conduct a meta-evaluation of reference-based automatic metrics for English sentence simplification using high-quality, human-annotated dataset, NEWSELA-LIKERT. We study the behavior of several evaluation metrics at sentence level across four different sentence simplification models. All the models were trained on the NEWSELA-AUTO dataset. The correlation between the metrics’ scores and human judgements was analyzed and the results used to recommend the most appropriate metrics for this task.

This article describes a novel framework for the detection of causal links between financial news and the subsequent movements of the stock market. The approach builds on and substantially improves a previously published in-house design for the detection and measurement of correlation between news and time series in the financial domain, which has been used here to detect a predictive causality relationship from news to prices and volumes of trade. While the original framework makes use of matrices of pairwise distances between companies, one based on news, the other - on financial performance, in order to produce a single measure of correlation between these two types of information for all traded companies, this article shows how the company contributing the most to the news-to-price/volume causal link can be singled out. The potential benefits of such information are made clear through its use in a straight-forward trading strategy, the results of which compare favourably to two strong, real-life alternatives that only make use of the time series.

Adaptive Operator Selection (AOS) is an approach that controls discrete parameters of an Evolutionary Algorithm (EA) during the run. In this paper, we propose an AOS method based on Double Deep Q-Learning (DDQN), a Deep Reinforcement Learning method, to control the mutation strategies of Differential Evolution (DE). The application of DDQN to DE requires two phases. First, a neural network is trained offline by collecting data about the DE state and the benefit (reward) of applying each mutation strategy during multiple runs of DE tackling benchmark functions. We define the DE state as the combination of 99 different features and we analyze three alternative reward functions. Second, when DDQN is applied as a parameter controller within DE to a different test set of benchmark functions, DDQN uses the trained neural network to predict which mutation strategy should be applied to each parent at each generation according to the DE state. Benchmark functions for training and testing are taken from the CEC2005 benchmark with dimensions 10 and 30. We compare the results of the proposed DE-DDQN algorithm to several baseline DE algorithms using no online selection, random selection and other AOS methods, and also to the two winners of the CEC2005 competition. The results show that DE-DDQN outperforms the non-adaptive methods for all functions in the test set; while its results are comparable with the last two algorithms.

Building a proactive and unobtrusive recom- mender system is still a challenging task. In the real world, buyers may be offered a lot of choices while trying to choose the item that best suits their preference. Such items may have many attributes, which can complicate the process. The classic approach in decision support systems - to put weights on the importance of each attribute - is not always helpful here. For instance, there are cases when users find it is hard to formulate their priorities explicitly. In this paper, we promote the use of pairwise comparisons, which allow the user preferences to be inferred rather than spell out. Our system aims to learn from a limited number of examples and using clustering to guide the selection of pairs for annotation. The approach is demonstrated in the case of purchasing a used car using a large, real-world data set.

In time series analysis research, there is a strong interest in discrete representations of real valued data streams. One approach still considered state-of-the-art is the Symbolic Aggregate Approximation (SAX) algorithm. The interest of this paper concerns the SAX assumption of data being highly Gaussian and the use of the standard normal curve to choose partitions to discretize the data. The SAX approach chooses partitions on the standard normal curve that would produce an equal probability for each symbol. This procedure is generally valid as a time series is normalized to have μ = 0 and σ = 1. However, there exists a caveat to this assumption of equi-probability due to the intermediate step of Piecewise Aggregate Approximation (PAA). We show in this paper that when PAA is applied, the distribution of the data is altered, resulting in a shrinking standard deviation that is proportional to the number of points used to create a segment of the PAA representation and the degree of auto-correlation within the series. Data that exhibits statistically significant auto-correlation is less affected by this shrinking distribution. As the standard deviation of the data contracts, the mean remains the same, however the distribution is no longer standard normal and therefore the partitions based on the standard normal curve are no longer valid for the assumption of equal probability.

This paper introduces a novel forecasting algorithm that is a blend of micro and macro modelling perspectives when using Artificial Intelligence (AI) techniques. The micro component concerns the fine-tuning of technical indicators with population based optimization algorithms. This entails learning a set of parameters that optimize some economically desirable fitness function as to create a dynamic signal processor which adapts to changing market environments. The macro component concerns combining the heterogeneous set of signals produced from a population of optimized technical indicators. The combined signal is derived from a Learning Classifier System (LCS) framework that combines population based optimization and reinforcement learning (RL). This research is motivated by two factors, that of non-stationarity and cyclical profitability (as implied by the adaptive market hypothesis [10]). These two properties are not necessarily in contradiction but they do highlight the need for adaptation and creation of new models, while synchronously being able to consult others which were previously effective. The results demonstrate that the proposed system is effective at combining the signals into a coherent profitable trading system but that the performance of the system is bounded by the quality of the solutions in the population.

This study analyzes two implications of the Adaptive Market Hypothesis: variable efficiency and cyclical profitability. These implications are, inter alia, in conflict with the Efficient Market Hypothesis. Variable efficiency has been a popular topic amongst econometric researchers, where a variety of studies have shown that variable efficiency does exist in financial markets based on the metrics utilized. To determine if non-linear dependence increases the accuracy of supervised trading models a GARCH process is simulated and using a sliding window approach the series is tested for non-linear dependence. The results clearly demonstrate that during sub-periods where non-linear dependence is detected the algorithms experience a statistically significant increase in classification accuracy. As for the cyclical profitability of trading rules, the assumption that effectiveness waxes and wanes with the current market environment, is tested using a popular technical indicator, Bollinger Bands (BB), that are converted from static to dynamic using particle swarm optimization (PSO). For a given time period the parameters of the BB are fitted to optimize profitability and then tested in several out-of-sample time periods. The results indicate that on average a particular optimized BB is profitable, active and able to outperform the market index up to 35% of the time. These results clearly indicate the cyclical nature of the effectiveness of a particular trading model and that a technical indicator derived from historical prices can be profitable outside of its training period.

This study investigates the characteristic of non-stationarity in a financial time-series and its effect on the learning process for Artificial Neural Networks (ANN). It is motivated by previous work where it was shown that non-stationarity is not static within a financial time series but quite variable in nature. Initially unit-root tests were performed to isolate segments that were stationary or non-stationary at a pre-determined significance level and then various tests were conducted based on forecasting accuracy. The hypothesis of this research is that when using the de-trended/original observations from the time series the trend/level stationary segments should produce lower error measures and when the series are differenced the difference stationary (non-stationary) segments should have lower error. The results to date reveal that the effects of variable stationarity on learning with ANNs are a function of forecasting time-horizon, strength of the linear-time trend, sample size and persistence of the stationary process.

Current approaches to instruction cache analysis for determining worst-case execution time rely on building a mathematical model of the cache that tracks its contents at all points in the program. This requires perfect knowledge of the functional behaviour of the cache and may result in extreme complexity and pessimism if many alternative paths through code sections are possible. To overcome these issues, this paper proposes a new hybrid approach in which information obtained from program traces is used to automate the construction of a model of how the cache is used. The resulting model involves the learning of a Bayesian network that predicts which instructions result in cache misses as a function of previously taken paths. The model can then be utilised to predict cache misses for previously unseen inputs and paths. The accuracy of this learned model is assessed against real benchmarks and an established statistical approach to illustrate its benefits.

This study analyzes the effectiveness of an Artificial Immune System (AIS) to model and predict the movements of the stock market. To aid in this research the AIS models are compared with a k-Nearest Neighbors (kNN) algorithm, an artificial neural network (ANN) and a benchmark market portfolio to compare simulated trading results. The analysis shows that the AIS produced overall accuracy results of 67% over a 20 year test period and that the increased complexity of the model was warranted by the statistically significant superior results when compared to the simpler instance-based approach of kNN. The accuracy results were comparable to those obtained from training the ANN and the trading results outperformed the market benchmark, providing evidence that the stock market had a degree of predictability during the time period of 1989-2008. In general the practice of using the natural immune system to inspire a learning algorithm has been established as a viable alternative to modeling the stock market when implementing a supervised learning approach.

Determination of accurate estimates for the Worst-Case Execution Time of a program is essential for guaranteeing the correct temporal behavior of any Real-Time System. Of particular importance is tightly bounding the number of iterations of loops in the program or excessive undue pessimism can result. This paper presents a novel approach to determining the number of iterations of a loop for such analysis. Program traces are collected and analyzed allowing the number of loop executions to be parametrically determined safely and precisely under certain conditions. The approach is mathematically proved to be safe and its practicality is demonstrated on a series of benchmarks.

Static analysis can be used to determine safe estimates of Worst Case Execution Time. However, overestimation of the number of loop iterations, particularly in nested loops, can result in substantial pessimism in the overall estimate. This paper presents a method of determining exact parametric values of the number of loop iterations for a particular class of arbitrarily deeply nested loops. It is proven that values are guaranteed to be correct using information obtainable from a finite and quantifiable number of program traces. Using the results of this proof, a tool is constructed and its scalability assessed.

Most software engineering methods require some form of model populated with appropriate information. Real-time systems are no exception. A significant issue is that the information needed is not always freely available and derived it using manual methods is costly in terms of time and money. Previous work showed how machine learning information derived during software testing can be used to derive loop bounds as part of the Worst-Case Execution Time analysis problem. In this paper we build on this work by investigating the issue of branch prediction.

This paper reports the results of a study of a specific type of concurrency in the Ant Colony System (ACS) algorithm. Studies of Cellular Automata (CA) have shown that the update mechanism used can have a dramatic influence on the dynamics of the CA. ACS is usually implemented with a sequential update mechanism. A new method for controlling the concurrency in a nature-inspired algorithm is introduced. Comprehensive tests on a wide range of problem instances are reported. The study found that concurrency levels had no statistically significant effect on ACS performance. This result is interesting because it contradicts what has been observed in another form of nature-inspired algorithm, namely CAs.

Most development, verification and validation methods in software engineering require some form of model populated with appropriate information. Realtime systems are no exception. However a significant issue is that the information needed is not always available. Often this information is derived using manual methods, which is costly in terms of time and money. In this paper we show how techniques taken from other areas may provide more effective and efficient solutions. More specifically machine learning is applied to the problem of automatically deriving loop bounds. The paper shows how taking an approach based on machine learning allows a difficult problem to be addressed with relative ease.

This article suggests that the parser underlying human syntax may have originally evolved to assist navigation, a claim supported by computational simulations as well as evidence from neuroscience and psychology. We discuss two independent conjectures about the way in which navigation could have supported the emergence of this aspect of the human language faculty: firstly, by promoting the development of a parser; and secondly, by possibly providing a topic of discussion to which this parser could have been applied with minimum effort. The paper summarizes our previously published experiments and provides original results in support of the evolutionary advantages this type of communication can provide, compared with other foraging strategies. Another aspect studied in the experiments is the combination and range of environmental factors that make communication beneficial, focusing on the availability and volatility of resources. We suggest that the parser evolved for navigation might initially have been limited to handling regular languages, and describe a mechanism that may have created selective pressure for a context-free parser.