Improve performance and robustness of knowledge-based FUZZY LOGIC habitat models

Previous criticisms of knowledge-based fuzzy logic modelling have identified some of its limitations and revealed weaknesses regarding the development of fuzzy sets, the integration of expert knowledge, and the outcomes of different defuzzification processes. We show here how expert disagreement and fuzzy logic mechanisms associated with the rule development and combinations can positively or adversely affect model performance and the interpretation of results. We highlight how expert disagreement can induce uncertainty into model outputs when defining fuzzy sets and selecting a defuzzification method. We present a framework to account for sources of error and bias and improve the performance and robustness of knowledge-based fuzzy logic models. We recommend to 1) provide clear/unambiguous instructions on model development, processes and objectives, including the definition of input variables and fuzzy sets, 2) incorporate the disagreement among experts into the analysis, 3) increase the use of short rules and the OR operator to reduce complexity, and 4) improve model performance and robustness by using narrow fuzzy sets for extreme values of input variables to expand the universe of discourse adequately. Our framework is focused on fuzzy logic models but can be applied to all knowledge-based models that require expert judgment, including expert systems, decision trees and (fuzzy) Bayesian inference systems.