Knowledge Representation Methods in Crop Pest Expert Systems
Authors: Zhang Manna, Zhang Wu, Jin Xiu, Zhu Cheng, Song Yifan, Hong Xun, Li Mengjie (College of Information and Computer, Anhui Agricultural University, Hefei, Anhui 230036)
Abstract
The effectiveness of knowledge representation and reasoning methods for crop pests is the foundation for accurate decision-making in crop pest expert systems. This article summarizes the advantages and disadvantages of common traditional knowledge representation, modern knowledge representation, and hybrid knowledge representation methods used in crop pest expert systems, and analyzes and forecasts their development trends. Hybrid knowledge representation methods are widely applied as they can effectively overcome the limitations of single representation methods and fully leverage the strengths of various methods. Neural network knowledge representation shows promising prospects due to its speed, high accuracy, and flexibility.
Keywords
Crop; Pest; Knowledge Representation; Expert System
1 Introduction
The crop pest expert system is an intelligent computer program that integrates various knowledge representation technologies and reasoning strategies to provide diagnosis, prevention, and inquiry services for agricultural producers and users［1］.
In 1965, Feigenbaum developed the world’s first expert system, DENDRAL, at Stanford University in the United States［2］. After half a century of development, expert systems have been widely applied, and scholars both domestically and internationally have applied expert systems in the field of crop pest control, achieving significant economic and social benefits［2-4］. The University of Illinois developed the world’s first soybean pest expert system, PLAN/ds［5］. In 2008, scientists at the University of Illinois developed the corn borer pest prediction system, PLANT/cd. CARISTI et al. developed the EPINFORM expert system for predicting wheat diseases［5-7］. Chen Zhulu［8］ developed a rice pest diagnosis system based on a feature-based diagnostic mechanism, inferring the causes of rice pest outbreaks through comprehensive analysis of various feature attributes. Yang Linan et al.［9］ developed a sweet corn pest intelligent diagnosis system based on the Android system using production rules, achieving good generality. Wen Haojie et al.［10］ designed a web-based cucumber disease diagnosis system, achieving a high recognition rate.
The crop pest expert system predicts or provides early warnings based on the knowledge representation and reasoning of crop pests, relying on knowledge representation and reasoning methods. Knowledge representation methods significantly affect the implementation form, efficiency, and resource consumption of knowledge processing. Therefore, when addressing a specific issue, effective knowledge representation is essential for the expert system to utilize this knowledge for reasoning and making effective decisions. This article categorizes the knowledge representation methods in crop pest expert systems into traditional knowledge representation methods, modern knowledge representation methods, and hybrid knowledge representation methods, analyzing the characteristics of each type of knowledge representation method and their applications in crop pest systems, while forecasting future development trends.
1 Traditional Knowledge Representation Methods in Crop Pest Systems
1.1 Logical Representation
Logical representation expresses the relationship between the subject and object of actions in predicate form, serving as a narrative knowledge representation method. Logical formulas can describe objects, properties, conditions, and relationships. There are two main types of logical representation: propositional logic and predicate logic.
Predicate logic representation is one of the earliest knowledge representation methods used in artificial intelligence. For example, Ouyang Da et al.［11］ used predicate logic to represent knowledge about wheat scab disease when establishing a prediction expert system for wheat scab in Shanghai in 1991, describing the knowledge “Wheat is infected with scab” as: infects(wheat, scab). Here, “infects” is the predicate representing the action; “wheat” is the subject, indicating the entity performing the action; and “scab” is the object, indicating the target of the action. The prediction accuracy of the system using predicate logic representation was not high［11-12］. Although predicate logic representation is natural, flexible, and clearly structured, it is difficult to represent uncertain knowledge and heuristic knowledge, and the reasoning process is lengthy and inefficient. Therefore, it is rarely used in modern crop pest systems［13］.
1.2 Production Rule Representation
Production rule representation, also known as rule representation, is expressed in IF-THEN format, where the IF clause describes the prerequisites of the rule, and the THEN clause states the conclusion of the rule. Production rule representation is primarily used to describe knowledge and state the control and interaction mechanisms between various process knowledge.
Due to its natural simplicity, clarity, flexibility, and ease of understanding, production rule representation is widely applied in the field of agricultural pests. For example, Zhang Yunfei［14］ established a pest knowledge base based on production rule representation and implemented pest diagnosis through forward reasoning; Hou Baohua［15］ used production rule representation to express barley pest knowledge, establishing a barley pest control system due to the fuzziness, complexity, and uncertainty of barley pest characteristics; Jia Qian［16］ used production rule representation in the design of a tomato greenhouse environment management system to establish a knowledge base for pest control, achieving practical utility. CHEN［17］ developed the CUPTEX/Seedking expert system for cucumber pest control based on production rule representation. Zuo Zhiping et al.［18］ proposed that using production rule representation for pest knowledge is beneficial for optimizing subsequent problem-solving.
Despite the many advantages of production rule representation, it cannot represent structural knowledge, and solving complex problems can lead to combinatorial explosion. Additionally, the solving process involves repeated matching, conflict resolution, and execution, resulting in low efficiency［19］. To address these issues, many researchers have proposed improvement methods.
1.3 Improved Production Rule Representation
Analyzing the basic form of production rule representation, production rules can be transformed into reasoning rules and computational rules. To address the handling of fuzzy knowledge, credibility factors r can also be introduced into computational rules［20］.
Reasoning rule: if (premise 1, premise 2, ⋯, premise n) then conclusion
Computational rule: conclusion -> value = FUNC[(premise 1 -> value, premise 2 -> value, ⋯, premise n -> value), r] (0 ≤ r ≤ 1).
This enhanced production rule system corresponds each reasoning rule to a computational rule, with the function FUNC varying alongside the reasoning rules. This method not only retains the advantages of traditional production rule representation but also improves reasoning accuracy. When the credibility r ≠ 1, flexible fuzzy reasoning can be achieved. For instance, Guo Xiaoyan et al.［21］ incorporated a confidence factor to describe the credibility of fuzzy knowledge when designing a corn pest system based on production rule representation, demonstrating broad application potential and value.
In addition to the aforementioned method of increasing the confidence factor, Deng Chao et al.［22］ proposed a “classified production rule” representation method to solve the combinatorial explosion problem of production rules. The essence of classification is to divide the conditions of production rules into five categories: sufficient, necessary, essential, important, and auxiliary based on different dependency relationships, and classify the rules accordingly. The “classified production rule” representation method limits the number of rules and effectively resolves the knowledge combinatorial explosion problem present in traditional production rules. The crop pest system designed using this method achieved an accuracy rate of up to 90% in practical tests.
1.4 Frame Representation
Frame representation, proposed by Minsky in 1975, is particularly adept at representing structural knowledge. It is a complex data structure that stores all knowledge about a specific event or object together, with a fixed subject representing a particular concept, object, or event, and lower layers consisting of slots (SLOT) representing attributes of each aspect of the subject［23］. The general form of frame representation is as follows:
The above example is from Cong Fei［24］, who used frame knowledge representation to establish a pest knowledge base in the development of a tomato pest diagnosis system. This system can provide timely and accurate diagnosis of pests that occur during tomato growth. Duan Yunpeng［25］ chose frame representation for the development of a corn pest and weed diagnosis expert system, as the symptoms and patterns of corn weed outbreaks are influenced by various natural conditions. The system demonstrated good performance in user interface, reliability, and system functionality.
Frame representation has strong expressive ability and a rich hierarchical structure, allowing for predictions about future situations based on past knowledge. This prediction closely aligns with human cognitive patterns, facilitating understanding［13］. However, a notable drawback of frame representation is its ineffectiveness in representing procedural knowledge and difficulty in generalizing new situations, often requiring combination with other methods for optimal performance.
1.5 Semantic Network Representation
A semantic network is a very effective knowledge representation method in artificial intelligence, expressed as a network graph through concepts and their semantic relationships. A semantic network describes the relationships between events, concepts, conditions, actions, and objects using nodes and labeled edges to form a directed graph, typically using AKO (a kind of) to represent relationships［26］. For example, Figure 1 shows the soybean pest knowledge represented by a semantic network.
Semantic networks are widely used due to their ability to flexibly and succinctly represent connections between nodes. For instance, Dong Changbo［27］ utilized semantic networks to represent soybean pest knowledge when developing a comprehensive service platform for soybean information, significantly improving the system’s decision-making speed and accuracy. Zhang Rong［28］ chose semantic network representation for rice pest knowledge, considering the complexity and broad coverage of the disciplines involved. However, the limitations of semantic networks primarily lie in their non-strictness in knowledge representation, leading to ambiguity, which restricts their widespread application in fields requiring precise knowledge representation.
2 Modern Knowledge Representation Methods in Crop Pest Expert Systems
2.1 Concept Maps
Concept maps, proposed by John F. Sowa in 1984, are a new knowledge representation method based on linguistics, psychology, and philosophy. They consist of finite, connected, directed bipartite graphs formed by concept nodes and relationship nodes. Each concept node corresponds to a specific or abstract concept, representing entities, attributes, states, and events; relationship nodes indicate the connecting relationships between concepts［29-30］. Concept maps are somewhat similar to semantic networks but differ in that they can naturally incorporate background knowledge, aligning more closely with human thought processes.
Concept maps can be represented in two forms: linear and display. The linear form is convenient for display and printing on computer terminals, while the display form is more intuitive and easier to understand. Below are the two representations of the knowledge “A cat sitting on a mat,” as shown in Figure 2.
Due to their simplicity, intuitiveness, and strong semantic expression capabilities, concept maps are widely applied in document clustering［31］, natural language processing［32］, text mining［33］, information retrieval［34-35］, and educational activities. However, the effective application of concept maps in crop pest expert systems has achieved limited results, making it a worthwhile research topic on how to effectively apply concept map knowledge representation in crop pest expert systems.
2.2 Object-Oriented Knowledge Representation Methods
With the in-depth development of object-oriented technology, object-oriented knowledge representation has gradually been applied in expert systems. Object-oriented knowledge representation uses objects as the basic unit to represent knowledge, combining various single knowledge representations such as production rules and frames based on object principles. Li Shuge［36］ proposed using object-oriented knowledge representation as a carrier for knowledge in a wheat pest diagnosis system, subsequently representing procedural knowledge as rule knowledge. This approach enhances the control of procedural knowledge through rules, ultimately combining rule-based and object-oriented knowledge into a new composite knowledge representation. Tests have shown that this object-oriented knowledge representation effectively resolves the issues of heterogeneous object reuse in expert systems.
While object-oriented knowledge representation has advantages such as encapsulation, modularity, inheritance, ease of maintenance, and scalability, it is not suitable for representing control and procedural knowledge［37］. Therefore, it must be combined with other knowledge representations to compensate for this drawback. Ding Weilong et al.［38］ employed an object-oriented “knowledge body, object block, component” knowledge representation method when constructing a greenhouse tomato cultivation management expert system. This method, based on a general framework production method, is simple and easy to learn, effectively representing various types of knowledge.
Using object-oriented representation for knowledge in crop pest systems facilitates the management, retrieval, and reuse of pest knowledge.
2.3 Neural Network Knowledge Representation Methods
Neural networks are abstract mathematical models reflecting the structure and function of the human brain. They consist of numerous individual neuron nodes connected by weighted values, with each node representing a computation called the activation function. Each connection between two nodes represents a weighted value for the signal passing through that connection. Complex networks are formed by interconnecting neuron nodes to simulate the human nervous system for knowledge representation, storage, and reasoning［39］. Neural network knowledge representation differs from production rule representation, which displays knowledge as a series of rules; instead, it is an implicit representation where relevant knowledge is “remembered” in the network through the weight matrix and system parameters. Neural networks are trained and learned by altering the connection weights of each neuron node in the network, acquiring knowledge through sample training and adapting to new environments. Once the network training is successful, it can output results in a short time［40］. Figure 3 illustrates a simple neural network, including three parts: input layer, hidden layer, and output layer. X1, X2, ⋯, Xn are input factors, and Y1, Y2, ⋯, Yn are output results, with Wij and Vjk representing weights.
With the rapid development of artificial intelligence, neural networks are widely applied in various fields due to their speed and effectiveness. In recent years, their application in crop pest systems has also increased. Jefferson developed the Soybug system based on neural networks for soybean pest control［41］. Jia Jiannan et al.［42］ trained 60 samples of two cucumber disease leaves using neural networks, testing 40 samples with a correct identification rate of 100%, indicating that using neural networks for identifying bacterial angular leaf spot and cucumber downy mildew is feasible. CHENG et al.［40］ proposed using convolutional neural network algorithms to identify and track pests in complex farmland environments. This algorithm effectively identifies camouflaged crop pests and achieves high recognition accuracy in high background noise, demonstrating significant practical value. However, BP neural networks face issues of low efficiency and susceptibility to local minima. To address these issues, some scholars have proposed improved BP neural networks. For example, Wang Xiaoxia et al.［43］ introduced additional momentum methods and adaptive learning rate methods into traditional BP neural networks, establishing a fuzzy BP neural network diagnostic model for diagnostic reasoning, exemplified by grape disease diagnosis, showing high diagnostic efficiency, practicality, generality, and flexibility.
Currently, neural network knowledge representation methods are widely applied in many areas due to their speed, effectiveness, accuracy, and high flexibility.
3 Hybrid Knowledge Representation Methods in Crop Pest Expert Systems
In the research process of crop pest expert systems, expert systems that implement a certain knowledge representation based on the characteristics of pest knowledge have achieved some application results. However, as research deepens, single knowledge representation methods struggle to effectively address all knowledge representations in the field. Therefore, hybrid knowledge representation methods have emerged, effectively combining mature knowledge representation methods from the past. Currently, well-established hybrid knowledge representation methods include the combination of predicate logic, production rules, and procedural methods; the combination of production rules, frames, and procedural methods; and the combination of semantic networks and production rules. The aforementioned frame representation and object-oriented representation methods involve hybrid knowledge representation. Yao Yuxia et al.［44］ established a maize weed knowledge base in a maize weed diagnosis expert system using a combination of object-oriented, frame, and production rule knowledge representations, primarily considering the accurate representation of maize weed knowledge. Hybrid knowledge representation methods can effectively overcome the limitations of single representation methods and fully leverage the strengths of various knowledge representation methods, gradually becoming the mainstream method of knowledge representation.
4 Conclusion
Through the study and analysis of different knowledge representation methods, it is evident that hybrid knowledge representation methods can effectively overcome the limitations of single representation methods and fully leverage the strengths of various methods, demonstrating broad application prospects. Neural network knowledge representation methods, due to their speed, high accuracy, and flexibility, will also become mainstream methods.
The processes of knowledge representation and reasoning are not isolated but complement each other. Therefore, finding the optimal matching pattern between knowledge representation methods and knowledge reasoning will be feasible and a topic worth exploring.
References omitted
Funding Projects: 2018 Anhui Provincial Key Research and Development Plan Project (1804a07020108); 2016 Agricultural Ministry Key Laboratory Open Fund Project (2016KL05); 2017 Anhui Provincial Major Science and Technology Special Project (17030701049); 2019 Anhui Provincial Key Research and Development Plan Major Project (201904a0620056)
Source: Zhang Manna, Zhang Wu, Jin Xiu, et al. Knowledge Representation Methods in Crop Pest Expert Systems [J]. Journal of Jianghan University (Natural Science Edition), 2019, 47(4): 378-384.
DOI: 10.16389/j.cnki.cn42-1737/n.2019.04.015
