Interactive Genetic Algorithms For Watershed Planning PDF Download

Degradation of watersheds is a major concern in areas where adverse climate effects and unsustainable use of the natural resources have caused extensive stresses to watershed systems (e.g., increased floods, increased droughts, worsened in-stream water quality) through the years. While considerable efforts are being made to generate technical solutions that focus on plans of spatially-distributed conservation practices (e.g., Wetlands, Filter Strips, Grassed Waterways, Crop Management practices, etc.) for restoration of existing conditions in the watersheds, adoption and implementation of these solutions require a better understanding of constraints faced by affected stakeholders and decision makers. Participatory modeling and design approaches have, as a result, become popular in the recent past to support a community's engagement during the modeling process and during development of potential scenarios of plans (or, design alternatives). And now, with new and ongoing developments in Web 2.0 technologies, there is an even greater need for research that examines how large number of stakeholders can be engaged in the development of design alternatives via the internet-based, decision support environments. The overarching goal of this research is to investigate how stakeholder participation ("humans") and Interactive Genetic Algorithms ("computer") can be coupled in a web-based watershed decision support system (DSS) called WRESTORE (Watershed REstoration using Spatio Temporal Optimization of REsources- http://wrestore.iupui.edu/), in order to generate user-preferred design alternatives of distributed conservation practices on a watershed landscape. An important component of this goal is to also improve the understanding of how human behavior on the graphical user interface (GUI) of the DSS can be observed and evaluated in real-time, and then learned from to further improve the performance of the underlying search algorithm. Four specific objectives were addressed in this work to accomplish the overall goal: · Objective 1: Observe interactions of multiple users with the GUI of a web-based watershed DSS (WRESTORE, http://wrestore.iupui.edu/) during interactive search experiments, and then use Usability metrics (response times, clicking events and confidence levels) to evaluate the differences and similarities in user behaviors and interactions. · Objective 2: Examine relationships between the type of users (e.g., stakeholders versus surrogates), the Usability metrics, and patterns in the watershed-scale plans of conservation practices generated by the multi-objective Interactive Genetic Algorithm embedded in WRESTORE. · Objective 3: Examine relationships between the type of users, the Usability metrics, and patterns in the user-preferred, sub-basin-scale plans of conservation practices generated by the multi-objective Interactive Genetic Algorithm embedded in WRESTORE. · Objective 4: Develop and test novel human-guided search operators that adaptively learn for patterns in user-preferred alternatives generated by the multi-objective Interactive Genetic Algorithm, and, as a result, improve the convergence rate of the search algorithm for generating design alternatives that conserve these learned patterns. Results show that there is a clear difference on how different types of users interact with the Interactive Optimization system. The observed relationship between confidence levels, time spent on a task, and number of mouse clicking events, indicated that participants who were able to use the WRESTORE GUI to gather more information and had a higher rate of time per number of clicks, tended to increase their levels of self-confidence in their own feedback. Also, when engaging with watershed stakeholders versus non-stakeholders (or, surrogates), 67% of the stakeholder participants steadily increased their average self-confidence levels as they continued to interact with the tool, in contrast to only 29% of surrogate participants who also showed an increase in their self-confidence levels through time. Such usability and confidence level evaluations provide assessments on which participant was potentially generating reliable feedback data for the search algorithm to use. An analysis of design alternatives generated by the individuals in both stakeholder and non-stakeholder groups showed that a majority (67%) of the stakeholder participants found a higher percentage (on and average 52%) of preferred design alternatives via the interactive search process. Also, users who were focused on assessing the suitability of design alternatives for the entire watershed trended to demonstrate a bias for one of the watershed-scale objective functions. In contrast, users, who were focused on assessing the suitability of design alternatives at only a few local sub-basins in the watershed, did not demonstrate any clear bias for any one of the watershed-scale objective functions. Additionally, patterns were observed in the design of decision alternatives generated by the human-centered search process, which further divulged potential user preferences related to the decision space for example, whether a specific participant preferred a certain practice over another, or a certain location over another for a specific practice. Finally, to improve the convergence rates of the Interactive Genetic Algorithm in WRESTORE, we investigated whether observed patterns in decisions (especially, when users were focused on local sub-regions of the watershed) can be used to improve the search for user-desire designs. A novel Interactive Genetic Algorithm with adaptive, human-guided, selection, crossover and mutation operators was proposed. The new algorithm was tested with six types of simulated participants (three deterministic and three probabilistic users) developed from the feedback data of three real participants. Results of search experiments with the novel adaptive IGA operators indicated a faster convergence than the default IGA, for two out of three deterministic simulated users. However, none of the probabilistic user showed a convergence different than the default values. This indicates that while current results indicate promise, there is need for additional research on adaptive, human-guided IGA operators, especially when noisy/stochastic users participate in the search. Additionally, adaptation of search operators have the potential to improve convergence rates when participatory design is done via Interactive Genetic Algorithms.