| Summary |
Chun-Yu Chen，2007.06
Feng Chia University
The liabilities of different parties involved in road traffic accidents need to be examined and appraised by fair and unbiased institutions, which are called Local Appraisal Committees (LACs) and Re-appraisal Committees (RACs) at two different levels of jurisdictions in Taiwan. However, several problems in current accident appraisal system can be identified. First, the committee members are usually overloaded by the requirement to handle thousands of accidents queued for appraisal every year. Also, because of the professional knowledge and expertise needed for conducting accident appraisals, very few qualified experts can be found. Second, there is often a lack of an effective and efficient mechanism for passing on the accumulated experience of senior appraisal experts when their service terms expire or the committee members are reorganized, which happens periodically. Third, for even very similar accident cases the appraisal results reported by different committees can be quite diverse. It has become an important and challenging issue to develop an effective expert system that can help the accident appraisal committees to enhance the consistency and efficiency of their decision-making process. Based on this, many accident appraisal expert systems have been developed, which are modeled by statistical methods, decision tree, and artificial neural network, respectively. However, some of them only proposed conceptual frameworks; some of them exhibits low correctness rate or performs under a black-box manner, leaving much room for improvement. Due to the powerful application flexibility and knowledge retrieval of logic inference rules, this study aims to develop and validate a logic rule-based expert system for accident appraisal. In order to retrieve representative logic rules from historical appraisal cases, genetic algorithms (GAs) are employed to select the best combination of logic rules (rules recovery). This expert system is expected to emulate the judging behaviors of committee members so as to shorten the appraisal time, provide experience inheritance, and produce consistent appraisal results.
To this end, three genetic logic rule-based models, Model 1-3, are respectively developed, each of which employs one gene to represent a corresponding state/control variable. Model 1 and Model 2 are party-based models which separately consider each of the parties involved in accidents; while Model 3 is case-based model which can simultaneously take two-parties involved in an accident into account. Moreover, Model 1 is a chromosome-based model which uses a chromosome to represent the combination of selected rules; while Model 2 and Model 3 are population-based models which use a chromosome to represent a selected logic rule and survived population to stand for the combination of selected rules. These three models aim to maximize the correctness rate, minimize the number of redundant rules and minimize the contradictory predicted results from selected rules. Noteworthy, because of the extremely lengthy chromosome, the model using a chromosome to stand for the selected rules with consideration of both parties involved (i.e. case-based) is not developed.
To investigate the performances of the proposed models and the interpretation of the corresponding recovery rules, the proposed models are developed based on a total of 538 two-car crash accident appraisal cases with 1076 parties involved, from 2000 to 2002 in Taiwan. By table of contingent, a total of 12 key variables are selected as potential state variables, which are significantly correlated to the degrees of liability assessed. The control variable is set as the degree of liability. All cases are randomly divided into a training set with a total of 377 cases (754 parties) and a validation set with a total of 161 cases (322 parties). Besides, comparisons to the models of artificial neural network (ANN) and discrimination analysis (DA) are conducted, which are also developed and validated by same data set. Results show that Model 2 performs better than Model 1 and Model 3 with the training and validation correctness rates of 77.87% and 70.19%, respectively, which are also much better than that of DA. Comparing to that of ANN, the training correctness rate of Model 2 is superior to that of ANN (77.19%), while the validation correctness rate of Model 2 is slightly lower than that of ANN (72.67%). The performances of the proposed models have been validated. Moreover, by examining the selected rules by Model 2, four state variables -- right-of-way, alcohol use, speeding, and type of road, frequently appear in the selected rules, suggesting the importance of these variables. |
