|
| Factors considered | Proposed BRP method | Existing wrapper methods |
| Bagging | Boosting | Feature section |
|
| Bridge Deck condition Assessment Measure | Condition rating from 0 to 9 from the poorest condition to near-perfect condition | Health index with 0–100% from worst to best condition | Health index with 0–100% from the worst to the best condition | Health index with 0–100% from the worst to the best condition |
| Basic Concept of the Method | Split or partition each node into two child nodes iteratively until reaching convergence | Generate multiple trees to get tree predictors and use them to achieve an aggregate predictor | Use DTA as the learning algorithm to construct different predictors by changing the distribution | Specify an appropriate subset of features by searching different subsets to improve prediction accuracy |
| Sample Size | Approximate 5,500 bridges with 25 years detailed information | 222 bridges from KDOT bridge database | 222 bridges from KDOT database | 222 bridges from KDOT database |
| Prediction Accuracy | 62–92% for 4 classes | 63%–74% | 63%–73% | 65%–75% |
| 84–95% for 2 classes |
| Ease of Use | Consists of four basic analytical steps: tree building, stopping tree building, tree pruning and optimal tree selection. Fully automatic and straightforward to use | Different random trees need to be built to get predictors and prediction accuracy. Deck health index needs to be manually computed | Multiple runs required to achieve the most appropriate weight for each instance. Deck health index needs to be manually computed | Multiple subsets of features need to be evaluated to get the optimal subset. Deck health index needs to be manually computed. Some results may be misleading |
|
