Krzysztof Michalak

Feature selection in corporate credit rating prediction

Credit rating assessment is a complicated process in which many parameters describing a company are taken into consideration and a grade is assigned, which represents the reliability of a potential client. Such assessment is expensive, because domain experts have to be employed to perform the rating. One way of lowering the costs of performing the rating is to use an automated rating procedure. In this paper, we assess several automatic classification methods for credit rating assessment. The methods presented in this paper follow a well-known paradigm of supervised machine learning, where they are first trained on a dataset representing companies with a known credibility, and then applied to companies with unknown credibility. We employed a procedure of feature selection that improved the accuracy of the ratings obtained as a result of classification. In addition, feature selection reduced the number of parameters describing a company that have to be known before the automatic rating can be performed. Wrappers performed better than filters for both US and European datasets. However, better classification performance was achieved at a cost of additional computational time. Our results also suggest that US rating methodology prefers the size of companies and market value ratios, whereas the European methodology relies more on profitability and leverage ratios.

Simheuristics for the Multiobjective Nondeterministic Firefighter Problem in a Time-Constrained Setting

The firefighter problem (FFP) is a combinatorial problem requiring the allocation of ‘firefighters’ to nodes in a graph in order to protect the nodes from fire (or other threat) spreading along the edges. In the original formulation the problem is deterministic: fire spreads from burning nodes to adjacent, unprotected nodes with certainty.

The Sim-EA Algorithm with Operator Autoadaptation for the Multiobjective Firefighter Problem

The firefighter problem is a graph-based optimization problem that can be used for modelling the spread of fires, and also for studying the dynamics of epidemics. Recently, this problem gained interest from the softcomputing research community and papers were published on applications of ant colony optimization and evolutionary algorithms to this problem. Also, the multiobjective version of the problem was formulated.

Auto-adaptation of Genetic Operators for Multi-objective Optimization in the Firefighter Problem

In the firefighter problem the spread of fire is modelled on an undirected graph. The goal is to find such an assignment of firefighters to the nodes of the graph that they save as large part of the graph as possible.

Infeasibility driven evolutionary algorithm with ARIMA-based prediction mechanism

This paper proposes an improvement of evolutionary algorithms for dynamic objective functions with a prediction mechanism based on the Autoregressive Integrated Moving Average (ARIMA) model. It extends the Infeasibility Driven Evolutionary Algorithm (IDEA) that maintains a population of feasible and infeasible solutions in order to react on changing objectives faster. Combining IDEA with ARIMA leads to a more efficient evolutionary algorithm that reacts faster to the changing objectives which profits from using information coming from the prediction mechanism and remains one time instant ahead of the original algorithm. Preliminary experiments performed on popular benchmark problems confirm that the IDEA-ARIMA outperforms the original IDEA algorithm in many cases.

The effects of asymmetric neighborhood assignment in the MOEA/D algorithm

Graphical abstractDisplay Omitted HighlightsThe MOEA/D algorithm is one of the leading multiobjective evolutionary algorithms.The paper describes the effects that cause the MOEA/D algorithm to converge asymmetrically.This is undesirable, because a multiobjective optimizer should explore the solution space evenly.The paper gives some guidelines on how to avoid such asymmetries. The Multiobjective Evolutionary Algorithm Based on Decomposition (MOEA/D) is a very efficient multiobjective evolutionary algorithm introduced in recent years. This algorithm works by decomposing a multiobjective optimization problem to many scalar optimization problems and by assigning each specimen in the population to a specific subproblem. The MOEA/D algorithm transfers information between specimens assigned to the subproblems using a neighborhood relation.In this paper it is shown that parameter settings commonly used in the literature cause an asymmetric neighbor assignment which in turn affects the selective pressure and consequently causes the population to converge asymmetrically. The paper contains theoretical explanation of how this bias is caused as well as an experimental verification. The described effect is undesirable, because a multiobjective optimizer should not introduce asymmetries not present in the optimization problem. The paper gives some guidelines on how to avoid such artificial asymmetries.

Evolutionary algorithm with a directional local search for multiobjective optimization in combinatorial problems

Evolutionary algorithms (EAs) are often employed to multiobjective optimization, because they process an entire population of solutions which can be used as an approximation of the Pareto front of the tackled problem. It is a common practice to couple local search with evolutionary algorithms, especially in the context of combinatorial optimization. In this paper a new local search method is proposed that utilizes the knowledge concerning promising search directions. The proposed method can be used as a general framework and combined with many methods of iterating over a neighbourhood of an initial solution as well as various decomposition approaches. In the experiments the proposed local search method was used with an EA and tested on 2-, 3- and 4-objective versions of two well-known combinatorial optimization problems: the travelling salesman problem (TSP) and the quadratic assignment problem (QAP). For comparison two well-known local search methods, one based on Pareto dominance and the other based on decomposition, were used with the same EA. The results show that the EA coupled with the directional local search yields better results than the same EA coupled with any of the two reference methods on both the TSP and QAP problems.

Sim-EA: An Evolutionary Algorithm Based on Problem Similarity

In this paper a new evolutionary algorithm Sim-EA is presented. This algorithm is designed to tackle several instances of an optimization problem at once based on an assumption that it might be beneficial to share information between solutions of similar instances. The Sim-EA algorithm utilizes the concept of multipopulation optimization. Each subpopulation is assigned to solve one of the instances which are similar to each other. Problem instance similarity is expressed numerically and the value representing similarity of any pair of instances is used for controlling specimen migration between subpopulations tackling these two particular instances.

ED-LS A heuristic local search for the multiobjective Firefighter Problem

The ED-LS method prioritizes defending those nodes that in the current graph state St have:1.a lower number of edges separating them from fire E(St,v).2.a higher number of edges connecting to untouched nodes D(St,v) among nodes equidistant from fireThus, in the example, the priority of nodes is: a, b, c.Display Omitted The Firefighter Problem (FFP) concerns stopping a threat spreading in a graph.The paper presents a new local search method (ED-LS) for the multiobjective FFP.The ED-LS uses heuristics to reduce the computational cost.ED-LS is tested with the MOEA/D and NSGA-II algorithms as global searches.Experiments show that the proposed approach effectively improves the results. One of the approaches to combinatorial optimization is to use global search methods, such as evolutionary algorithms combined with local search procedures. Local search can be very effective in improving the quality of solutions, however, searching large neighbourhoods can be computationally expensive. In this paper a new local search method is described that is dedicated to the Firefighter Problem (FFP). The Firefighter Problem concerns protecting nodes of a graph in order to stop a threat that is spreading in the graph. Because the resources used for protection are limited, the goal is to find the best order in which the nodes are protected and thus the FFP is a combinatorial optimization problem. One of the key elements in the design of a local search method is the definition of the neighbourhoods of solutions from which the local search starts. The method proposed in this paper (ED-LS) aims at lowering the computational cost of the local search by reducing, based on certain heuristics, the size of these neighbourhoods.In this paper the ED-LS is compared to an optimization without the local search, a local search using non-reduced neighbourhoods and a method that reduces the neighbourhood by checking only some randomly selected neighbours. The results show that the ED-LS improves the performance of the algorithm with respect to no local search at all and to the local search using non-reduced neighbourhoods, supporting the view that reducing the size of the neighbourhood is beneficial. Also, the ED-LS produces better results than when the size of the neighbourhood is reduced randomly, which means that the proposed heuristics are indeed selecting promising neighbours effectively. In the last part of the experiments the ED-LS is further improved to obtain faster improvement of solutions at the beginning of the optimization run.

Improving Classification of Patterns in Ultra-High Frequency Time Series with Evolutionary Algorithms

This paper proposes a method of distinguishing stock market states, classifying them based on price variations of securities, and using an evolutionary algorithm for improving the quality of classification. The data represents buy/sell order queues obtained from rebuild order book, given as price-volume pairs. In order to put more emphasis on certain features before the classifier is used, we use a weighting scheme, further optimized by an evolutionary algorithm.