Expert systems of real time as key tendency of artificial intelligence in tax administration

1. Introduction

In difficult conditions of modeling, classical modeling techniques are ineffective or even completely unacceptable.

This is due to the fact that it is impossible to describe the reality adequately with the help of a small number of model parameters, because the calculation of a model requires too much time and computing resources, and, most importantly, conditions of these methods application are not fulfilled and, therefore, it becomes impossible to use appropriate statistical criteria to evaluate the adequacy of the produced models.

Because of the afore-named drawbacks of traditional methods analytical systems of a new type have been actively developing of the last ten years. They are based on the artificial intelligence technologies that mimic natural processes, such as the activity of neurons in the brain or the natural selection process.

Artificial Intelligence (AI) is an area of scientific knowledge, bringing together a large number of areas involved in the study of the principles and laws of human mental activity and modeling tasks that are traditionally considered to be intellectual (Biryukov, 2011).

According to modern presentation, artificial intelligence (AI) is defined "as a scientific discipline whose goal is to develop hardware and software, allowing the user - non-programmer to set and solve the tasks that are traditionally considered intellectual, communicating with computers on a limited subset of a natural language."

The step of expert systems (ES) inception in the "bowels" of AI systems is a shift from the paradigm of heuristics to the paradigm of developing ways by which a professional expert presents nonformalized knowledge, methods and informal rules, which he uses in making decisions.

 Expert System (ES) - a program or set of programs, that makes it possible to present and systematize the expert knowledge in a certain application area in a suitable computer form and, based on this view, to solve applied problems like an expert at the request of users (Biryukov, 2011).

ES relates to knowledge-based systems, and includes the characteristic elements of these systems:

• Knowledge Base (KB). In the self-learning systems, knowledge base contains information which is the result of previous problems solutions;

• a mechanism for obtaining solutions (inferential mechanism);

• means of interface to communicate with the user.

The essence of expert system structure is illustrated in Figure 1.

Designed to the date ES usually solve applied problems of the following classes:

• preparation of semantic descriptions of objects on the input data (e.g., interpretation of symbols, signals);

• prediction of certain disorders (e.g., symptoms of diseases);

• contingency planning;

• monitoring and others.

Figure1. The structure of the expert system

ESs are effectively used in various fields of human activity: medicine, geology, economics, and others. There are several reasons for this:

1) there is an opportunity to solve previously unavailable, badly formalized problems involving a new, specially designed for this purpose mathematical apparatus (semantic networks, frames, fuzzy logic, neural networks);

2) ESs are focused to be exploited by a wide range of professionals (end users), who communicate with the system in dialog using comprehensible reasoning techniques and the terminology of specific subject area;

3) the use of ES can dramatically increase the effectiveness of the solutions taken by ordinary users, due to the accumulation of knowledge in ESs, including the knowledge of the highest qualification experts.

When creating artificial intellectual systems a large number of information technologies (ES, advise-giving systems, decision-making support systems, execution of decisions) are used. The common feature of these techniques is the use of some form of human knowledge. If we highlight the technology aimed to solve economic problems, this class of systems will be called "economic advice-giving systems" (EAS).

Here is the definition of EAS taken from scientific literature: "Under the EAS we mean any software-based product that reflects the economic specialist professional knowledge, his skills and experience used in the process of issuing solutions for the user." Romanov & Odintsov (2000) worked out the classification of EAS, which is based on the principle of simulation of the expert’ thinking processes and divided EAS into two classes:

1) EAS reproducing conscious thought processes of a man;

2) EAS reproducing unconscious thought processes of a man.

Decision makers in real control systems in the economic field usually carry out the following thought procedures:

• draw conclusions and develop management decisions based on the analysis of complete, incomplete and unreliable knowledge, i. e. in the conditions of uncertainty;

• explain and can justify why they have come to this conclusion;

• improve their knowledge, re-systematize them, study their own and others’ experience;

• make exceptions to the rules; use contradictory and improbable information;

• determine the level of their competence, i.e. determine whether they can or cannot make decision in this particular case.

2. Methodology and Data

2.1 Review of neural network instruments for Mathematical Modeling

Neural network models are widely used in the ECs to represent and accumulate knowledge (McCulloch & Pitts, 1943; Haykin, 1999; Chernik, 2010; Yasnitsky, 2010). Note that there are three reasons for the rapid development of methods for neural network modeling in general and particularly in the sphere of EAS:

1) In the neural network models parallel calculation method is implemented, i.e. several steps for computing operations are carried out at the same time. Due to this, the speed of a neurocomputer (electronic structure or neyroemulyator) increases sharply.

2) Neural network model does not require prerequisites of classical regression analysis, which is particularly important for the study of economic systems, where these prerequisites may not be fulfilled.

3) Though neural network models are parametric, they do not require pre-guessing of the form (structure) for the model.

The use of neural networks provides the following useful properties for the model:

1) Non-linearity. Artificial neurons can be linear or nonlinear. Neural networks which are constructed from compounds of non-linear neurons are nonlinear themselves. Moreover, this nonlinearity is of a special kind, since it is distributed over the network. Non-linearity is an important feature, particularly if the mechanism, responsible for the formation of an input signal, is also non-linear.

2) Display of the input information into the output information. One of the most popular paradigms of learning in neural network system is training with the teacher. This implies a change of synaptic weights based on a set of labeled training examples. Each example consists of an input signal and the corresponding desired response. From many examples one is randomly selected, and the neural network modifies the synaptic weights to minimize the differences between the desired output signal, and the signal, formed by a network in accordance with the selected statistical test. This training is carried out as long as the changes in synaptic weights will be insignificant.

3) Adaptability. Neural networks have the ability to adapt their synaptic weights and the very structure of the model to the changes in the environment. In particular, a network trained to operate in a particular environment can easily be retrained to operate in the conditions of minor oscillations of the medium parameters in the environment. Moreover, for operation in nonstationary environments, the neural networks, that change synaptic weights in real time, can be created. At the same time it should be noted that adaptability does not always lead to sustainability, for example, the adaptive system with parameters rapidly changing over time, can quickly react to extraneous field that cause the loss of productivity. In order to use all the advantages of adaptability, the basic parameters of the system must be sufficiently stable so to ignore external interference, and flexible enough to provide a response to significant changes in the environment.

An important restrictive feature of the neural networks used in EAS for representation knowledge is that unlike other models reproducing determined connections, clearly articulated by the expert, the neural network is not able to explain its results. Therefore in instrumental ES where it is possible to use several models of representation of knowledge, the neural network should be complemented by logic or production models.

2.2 Review of Models for Presentation of Knowledge

Consider the advantages and disadvantages of models for representation of knowledge in the EAS.

Table 1. Comparative characteristics of models for representation of knowledge in EAS

We give short comments to Table 1.

Logic models. The main idea of the approach in constructing logical models for knowledge representation is that all information necessary for solving applied problems, is considered as a collection of facts and statements that are presented as formulas in some kind of logic (Pospelov, 1989; Hayes-Roth & Jacobstein, 1994). This class of models is based on the predicate calculus and deals with the statements (affirmations), which can be true or false using Boolean algebra operation (disjunction , conjunction , negation , implication , and others.). Complex statements can be constructed in KB. For example,

where P, Q, R are predicates; g () - any logic function, for example, implication; x, y are object variables (objects).

The truth of this formula is determined by the specific values of the variables (x, y); the variables and the function values can take only two values: 1 and 0.

Frame models. In 1974 this type of models was first proposed by Marvin Minsky (1975, 1979), a professor at Massachusetts Institute of Technology (USA). A frame was considered as a minimal description of an object, i.e., a set of its features (attributes); the omission of any attribute makes the description of the frame impossible. Frame sample:

where Z is the frame name; Yi is the slot name; Ti – the slot value; Ai - the name of the associated procedure.

Slots comprise a frame basis. Slots are some unfilled structural elements of a frame. When filling the slot, this frame is associated with the considered situation, the object, the phenomenon. If a frame has empty slots, it is called a frame-prototype and a frame with filled slots is called a frame-example (or frame-instance).

Semantic networks. Existential graphs can be considered the progenitors of modern semantic networks. Existential graphs were proposed by Charles Pierce (1909). They are used to represent logical statements in the form of special diagrams. Pierce called this method “the logic of the future”. This is a chart pattern in the form of a graph, its vertices (nodes) are associated with some concepts (objects, events, processes), and the links (arcs) of the graph reflect the relationship between these concepts. The links can be diverse: temporal, spatial relations, cause and effect links, and others.

Production models. The term “products” was introduced by the American mathematician E. Potts in 1943. Nowadays, along with frames, products are the most popular forms of knowledge representation. The knowledge is presented in the form of sentences (production rules): “If A, then B”, where A, B are statements. In Logic this rule of deductive inference is called “Modus ponens”. For example: “If the temperature of a person is high, he should see a doctor” (Gavrilova & Khoroshevsky, 2000).

In general, a production rule is understood as the following expression:

Where i is the name of the product (its serial number or nomination); Q is the scope of the product application; P is the condition for the applicability of a product (predicate);  is “the core of a product” (A is a condition, B is an action);  is the sign of a sequence, which often coincides with a logical implication; N – product post condition, indicating what should be done after action B.

In practice, in order to describe quite complex objects (processes, events) not one product, but the set of products is used. When working with these systems of products, there may be questions about the systems’ completeness, consistency, the extension of the gained knowledge. The totality of these productions (rules) generates a "tree"-type structure of the arguments. In this structure the initial (root) vertex corresponds to the original statement, and the end (terminal) vertexes ("leaves") correspond to the results of reasoning. Each reasoning must contain a group of true statements corresponding to the intermediate tree nodes, situated on the path from the root to the definite terminal node. The search of this path can be performed by means of various algorithms, called inferential mechanism.

The abilities for effective work with the knowledge presented in the form of production models, are embedded in artificial intelligence languages LISP and PROLOG, which are specifically designed to treat symbolic information and automate logical reasoning (Vasilyev, 1981).

Fuzzy logic models. In 1964 L. A. Zadeh put the basis of fussy logic, i.e. a new approach to analyzing complex systems and decision-making processes under uncertainty (1996). The essence of this approach is as follows:

• fussy logic uses the so-called linguistic variables, that are e[pressed by whose terms (symbols) (words, phrases, sentences), expressed in natural language;

• simple relationships between the variables are described using fuzzy statements;

• complex relationship are described by the fuzzy algorithms for special operations on fuzzy sets (equivalence, inclusion, union, intersection, complement, the algebraic sum and product concentration, stretching, etc.). Fuzzy algorithm is an ordered set of fuzzy rules, formulated by vague indications (terms).

L. Zadeh’s ideas, focusing on the modeling of decision-making processes under uncertainty, have a lot of supporters and are widely used as a tool for the building of AI real systems.

The technology using fuzzy logic promotes the development of systems using intuition and engineering knowledge (know-how). Fuzzy logic uses the concepts of everyday language to determine the behavior of the system. It allows to create robust, fault-tolerant systems.

To sum up comments on Table 1, we note that the EAS can include a combination of different types of models for knowledge representation. This must generate the synergistic effect (the emergence) by strengthening the advantages of the basic model and reducing its negative characteristics and limitations. For example, when there is a combination of fuzzy and neural network models of knowledge representation in the EAS, fuzzy inference systems allow giving clear interpretation of the performed actions, but they cannot be taught, i.e. to perform automatic setting of parameters for membership functions on the basis of known information. In contrast, the neural network can customize their parameters (weights), but the functions, which they realize, can’t be clearly interpreted. The most effective way for hybridization of fuzzy logic and artificial neural networks, is a neuro-fuzzy system (more precisely - "neural inference system"), which, on the one hand, can be regarded as a fuzzy inference system (and thus, to interpret clearly the obtained results), and on the other hand - as an artificial neural network, that contain a special type of neurons and therefore, it can be trained.

Thus, fuzzy logic is a technology that enables the development of systems using intuition and engineering knowledge (know-how). Fuzzy logic uses the concepts of everyday language to determine the behavior of the system. It allows building robust, fault-tolerant systems (Zimmermann, 1996).

3. Results and Practice

Russian scientist Dmitry Chernik (2010) describes the ES, which is used by the Canadian Internal Revenue Service for verification of a company income tax and VAT. To develop the expertise rules for the selection of taxpayers in Canada, a group, consisting of 30 most qualified tax inspectors was created. These inspectors told the experts in artificial intelligence (AI), why some declarations seem suspicious to them, what things should be paid priority attention during the checking, and what amount of additional charges should be expected. To perform field tax audits, all these rules were introduced into the system of the computerized selection of taxpayers. The following sources of data are used: the data from tax returns, the data from previous field tax audits, the data on the structure of earnings in the area where the taxpayer lives. This ES allows to look through tax returns in automatic mode and to classify them into two classes: Class 1 – “field tax audits should be performed”; Class 2 – “it is not necessary to inspect”. The value of the expected additional charges is estimated in Class 1.

The US Federal Revenue Service uses the following combined (hybrid) model for selection of taxpayers for performing field tax audits (Chernik, 2010). A special data base of individuals and small and medium-sized businesses is collected according to the results of their special scrutiny, conducted in the framework of the Program "Measuring of taxpayers’ law-abiding." Randomly selected tax returns are classified by the main source of income. A discriminatory function is the amount of income (or gross income of the company). The classes’ labels are known from the results of previous checks. Then, using statistical methods specialists estimate the probability of additional accruals after the documentary check of the declaration from this class. This statistical model links the simulated core indicator with others, which the taxpayer showsin his declaration. This model is constructed as follows. Take a sample of declarations for the class, where the results of additional charges are known from previous audits and all these declarations are divided into two subclasses: 1 - "should have been checked"; 2 - " should not have been checked " (0 or 1).

The belonging of the declaration to one of these two categories is the modeled variable Y.

The simplest ES, which takes into account the uncertainty factor, is as follows. For simple regression:

a destructive version of multi-stage OLS is applied, i.e., first, a large number of order factors (150) is included in (4), then their number is reduced by an order and more, i.e. only the most significant factors for the explanation of the simulated variable are retained. Thus, the model of knowledge representation in the EC is a linear or non-linear regression equation (4), supplemented by the following production rule:

The required value interval is given by the expert.

In this regression equation, non-absolute values of these factors  from the declaration are used, but dimensionless complexes, formed from these factors.

After receiving the model (4), this simple regression formula is applied to all selected declarations of this class. The selected declarations are submitted to a highly qualified tax inspector (called a «classifier " in the US) for censorship. The classifier scans all the selected declarations and gives his verdict about each one: whether it should be checked or not, and what points should be especially focused on during the verification. In the United States, approximately half of all field inspections of taxes on personal and corporations income is organized using this ES selection.

According to D. Chernik’s study, tax services of many countries process tax returns using statistical methods such as regression and discriminant analysis, and according to the results of this analysis they build formulas, which allow to draw conclusions about new tax declarations: whether the check of a taxpayer promises large accrual or not. The documentary audit of taxpayers selected in such a way, is given priority.

4. Conclusions

Despite the intensive development of the theory and practice of neural network modeling in relation to the difficult conditions of economic systems modeling (hard formalization of the processes and interactions in the system), very noisy data (deliberate distortion of the tax base) because of the unknown laws of noise distribution, many problematic issues of neural network modeling of these systems have been either poorly researched or not researched at all:

1. The budgetary system of any level faces the problem of the insufficient amount of own funds for the projects focused on the end result and in the broader sense, for the budget system functioning, i.e. providing timely and quality services for budget spending units, ensuring the preservation of the financial and social stability and the development of the territories. Therefore many regions get subsidies. When transfers and subventions are allocated, the regional and municipal accounting documentation is distorted in the direction of increasing the budget deficit. In some cases budget planning is still based on the principle “from the achieved level”. So, one of the urgent problems is forecasting of budgets’ filling, particularly, in municipalities, taking into account the risk of going beyond the confidence limits of the forecast based on hybrid neural network models.

2. Lack of computer techniques for multi-criteria ranking of budgetary institutions and organizations, which would allow to assess objectively the results of the organization’ activities at the current time and a forecast period. Such techniques would allow to distribute transfers more equitably and efficiently in terms of governance and to determine the directions for the rational development of budgetary organizations.

3. The taxpayers’ activity is characterized by uncertain external and internal environment. The result of these trends is the spread of the output parameters for the organizations’ economic activity, which in many cases determines the high risk for inefficiency of tax audits. The objectives of the tax control are creative in nature; they require specialists with extensive knowledge, experience and developed intuition. Therefore, the transition to the mathematical formalization of the decision-making stages faces a number of difficulties associated with the problem of modeling for poorly formalized systems.

The Federal Tax Service of the Russian Federation use information technology for desk audits, selection of taxpayers for on-site inspections, These technologies are generally aimed to automate the monitoring of declared reporting data, their analysis for the logical consistency of the interrogation mode, where each subject of taxation is analyzed in turn.

The main disadvantage of the existing methods of tax control is as follows: all the technology of tax audits planning is subjective.

In addition, a number of factors increase the tax uncertainty, such as: the increasing complexity and diversification of the taxpayers’ activity, the diversity of the legal aspects of transactions between the taxpayers, the taxpayers’ risky activity, the raising stochasticity of the international competitive environment.

In such circumstances, there is need for new computer technologies for taxpayers’ selection.

So, we can conclude, that the level of development of theoretical and methodological foundations for neural network modeling in intelligent ECS for economic systems, does not meet the requirements of the practice, due to the ongoing process of reforming in the budget and tax system of the Russian Federation.

The modernization of the fiscal system in Russia and new financial and fiscal instruments pose new challenges requiring innovative solutions and rapid application in practice.

Solving this problem is aimed at solving algorithmically complex problems, as well as accumulation of scientifically based knowledge about the object, i.e., it is designed to maintain the existing system of economic models of the object of study and to add it with missing models and objectives. In applied economic aspect the problem is focused on improving the efficiency of state administration in the area of fiscal systems at all levels.

REFERENCES

1. Bestens D.-E., Van Den Berg V.-M., Wood D. (1998). Neural networks and financial markets: decision-making in trade. M.: TVP.

2. Biryukov, A. N. (2011). The theoretical basis for the development of neural network models in the system of tax administration. Ufa: RB Academy of Sciences, Guillem.

3. Biryukov, A. N. (2011). Bayesian regularization neural network models ranking and clustering of economic projects. Ufa: RB Academy of Sciences, Guillem.

4. Vasilyev, F. P. (1981). Methods for solving of extreme problems. Moscow: Nauka.

5. Gavrilova, T. A., & Khoroshevsky, V. F. (2000). Knowledge Databases for Intelligent Systems. SPb.: Peter.

6. Zade, L. (1996). The concept of a linguistic variable and its application to the adoption of approximate solutions. M.: Mir.

7. Minsky, M. L. (1979). Frames for knowledge representation. M.: Mir.

8. Minsky, M. L. (1975). A Framework for Representing Knowledge. In: Patrick Henry Winston (ed.), The Psychology of Computer Vision., New York (U.S.A.), McGraw-Hill.

9. McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. In K.E. Shannon & John McCarthy (Eds.), Machines (pp. 363-384). M.: Publishing House of Foreign Literature.

10. Melville, J. K. (1968). Charles Peirce and pragmatism. M.: Moscow State University.

11. Pospelov, D. (1989). "Guide to AI-2 ." Prospects for the development of computer technology. Kn.2. Intellectualization computer. Moscow.

12. Romanov, A. N., & Odintsov, B. E. (2000). Council Information systems in economy. M.: UNITY-DANA.

13. Hayes-Roth, F. & Jacobstein, N. (1994).The State of Knowledge-Based Systems. Communications of the ACM. Vol. 37, pp. 27-39.

14. Haykin, S. (1999). Neural Networks: A Comprehensive Foundation. (2nd ed.). New Jersey: Prentice-Hall.

15. Zimmermann, H. J. (1996). Fuzzy Set Theory and its Applications. (3rd ed.). Dordrecht: Kluwer Academic Publishers.

16. Chernik, D. G. (2010). Tax and Taxation. Moscow.

17. Endryu, A. M. (1985). Artificial Intelligence. M.: Mir.

18. Yasnitsky, L. N. (2010). Introduction to Artificial Intelligence. (3rd ed.). M.: Academy.