Boqin Feng

Automated Test Oracle Based on Neural Networks

In this paper an attempt has been made to explore the possibility of the usage of artificial neural networks as automated test oracle. Automated test oracle includes capabilities to generate expected output and compare it with actual output automatically. It is important for automated software testing. But there are very few techniques to implement it. In this paper, an insensitive oracle is proposed. It generates approximate output that is close to expected output. The actual output is then compared with the approximate output in an interval. The relation between inputs and outputs of an application under testing is described as a function. When it is a continue function, neural networks are used to estimate the output after training. By the method, automated oracle can be implemented and precision be adjusted by parameters. It can save a lot of time and labor in software testing

Neural Networks Based Test Cases Selection Strategy for GUI Testing

The purpose of graphical user interfaces (GUI) testing is to diagnose and expose faults in planning time. It is difficult because the input space of GUI is extremely large due to different permutations of inputs and events. To test GUI needs to run a lot of test cases. Neural networks (NN) were explored to reduce test cases to expose new faults. The main idea is as follows. Firstly, NN was trained by subset of test cases that had executed and their test results. Trained NN could recognize fault patterns that had been exposed. Secondly, from the test suite that hadnt been executed, trained NN was used to select test cases that dont belong to the fault patterns. The test cases selected were more likely to expose new faults in GUI. By the method new faults could be exposed by executing fewer test cases. The experimental results show that the strategy is effective

Compact Rule Learner on Weighted Fuzzy Approximation Spaces for Class Imbalanced and Hybrid Data

Rough set theory is an efficient tool for machine learning and knowledge acquisition. By introducing weightiness into a fuzzy approximation space, a new rule induction algorithm is proposed, which combines three types of uncertainty: weightiness, fuzziness and roughness. We first define the key concepts of block, minimal complex and local covering in a weighted fuzzy approximation space, then a weighted fuzzy approximation space based rule learner, and finally a weighted certainty factor for evaluating fuzzy classification rules. The time complexity of proposed rule learner is theoretically analyzed. Furthermore, in order to estimate the performance of the proposed method on class imbalanced and hybrid datasets, we compare our method with classical methods by conducting experiments on fifteen datasets. Comparative studies indicate that rule sets extracted by this method get a better performance on minority class than other approaches. It is therefore concluded that the proposed rule learner is an effective method for class imbalanced and hybrid data learning.

Multi-agent based multi-knowledge acquisition method for rough set

The key problem in knowledge acquisition algorithm is how to deal with large-scale datasets and extract small number of compact rules. In recent years, several approaches to distributed data mining have been developed, but only a few of them benefit rough set based knowledge acquisition methods. This paper is intended to combine multiagent technology into rough set based knowledge acquisition method.We briefly review the multi-knowledge acquisition algorithm, and propose a novel approach of distributed multi-knowledge acquisition method. Information system is decomposed into sub-systems by independent partition attribute set. Agent based knowledge acquisition tasks depend on universes of sub-systems, and the agent-oriented implementation is discussed. The main advantage of the method is that it is efficient on large-scale datasets and avoids generating excessive rules. Finally, the capabilities of our method are demonstrated on several datasets and results show that rules acquired are compact, having classification accuracy comparable to state-of-the-art methods.

On mining rules that involve inequalities from decision table

We introduce the notion of generating decision rules that involve inequalities. While a conventional decision rule expresses the trivial equality relations between attributes and values from the same or different objects, inequality rules express the non-equivalent relationships between attributes and values. The problem of mining inequality rules is formulated as a process of mining equality rules from a compensatory decision table. In order to mine high-order inequality rules, one can transform the original decision table to a high-order compensatory decision table, in which each new entity is a pair of objects. Any standard data-mining algorithm can then be used. We theoretically analyze the complexity of proposed models based on their meta-level representation in cognitive informatics. Mining inequalities in decision table makes a complementary feature of a rule induction system, which may result in generating a small number of short rules for domains where attributes have large number of values, and when majority of them are correlated with the same decision class.

CompactLEM2: A scalable rough set based knowledge acquisition method that generates small number of short rules

The complexity of knowledge plays an important role in the success of any types of knowledge acquisition algorithms performing on large-scale database. LERS (learning from examples based on rough sets) system is a rule based knowledge acquisition system that is characterized by excellent accuracy, but the complexity of generated rule set is not taken into account. This may cause interpretation problems for human and the classification knowledge may over fit training data. In this paper, CompactLEM2 is proposed as a scalable knowledge acquisition method that extracts rule set with easily understood rule forms, i.e., small size of rule set and short rule forms, without sacrificing classification accuracy. The main advantage of CompactLEM2 is its high efficiency. It can also produce compact rule set that fully or approximately describes classifications of given examples. We theoretically and experimentally show that CompactLEM2 exhibits log-linear asymptotic complexity with the number of training examples in most cases. We also present an example to illustrate characteristics of this algorithm. Finally, the capabilities of our method are demonstrated on eleven datasets. Experimental results are encouraging, and show that the length of extracted rule forms are short, and size of rule set is small, keeping the same level of classification accuracy of other rule acquisition methods in LERS system.

Using Wavelet Support Vector Machines to Generate Expected Outputs

A test case consists of a set of inputs and a list of expected outputs. To automatically generate the expected outputs for the test case is rather difficult. An approach based on wavelet support vector machines (WSVM) is proposed to overcome it. After training, WSVM is used to automatically generate the expected outputs, which approximate the correct outputs. Actual outputs of the application under testing (AUT) are then compared with expected outputs in an interval to determine if there is a failure. Experiment and comparison show that the method is effective and can save a lot of time and labor in software testing

Feature Based Rule Learner in Noisy Environment Using Neighbourhood Rough Set Model

From the perspective of cognitive informatics, cognition can be viewed as the acquisition of knowledge. In real-world applications, information systems usually contain some degree of noisy data. A new model proposed to deal with the hybrid-feature selection problem combines the neighbourhood approximation and variable precision rough set models. Then rule induction algorithm can learn from selected features in order to reduce the complexity of rule sets. Through proposed integration, the knowledge acquisition process becomes insensitive to the dimensionality of data with a pre-defined tolerance degree of noise and uncertainty for misclassification. When the authors apply the method to a Chinese diabetic diagnosis problem, the hybrid-attribute reduction method selected only five attributes from totally thirty-four measurements. Rule learner produced eight rules with average two attributes in the left part of an IF-THEN rule form, which is a manageable set of rules. The demonstrated experiment shows that the present approach is effective in handling real-world problems.