Date of Award

Spring 1-1-2012

Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Elizabeth Bradley

Second Advisor

Nichole Barger

Third Advisor

Nikolaus Correll

Fourth Advisor

Richard Byrd

Fifth Advisor

Manuel Laguna


In drought-prone environments, interactions between plants can enable individuals to conserve water. The physical location of each individual not only influences its resource needs, but also affects the availability of resources on the landscape. This thesis seeks to model this observed behavior in natural systems and use that model to create a new optimization approach for water conservation in residential systems in drought-prone climates. In this research, the arrangement of plants on a landscape is treated as a location optimization problem where the objective is to find the best locations for a given set of plants on a landscape with a given set of resources.

The biological properties of the domain make this optimization problem novel. The search space is influenced by what I call \textit{neighbor effects}, which include heterogeneity, locality, and feedback. Plants are heterogeneous, both in their requirements for growth and in their effects on their surroundings. Interactions with other plants are primarily local, and can either increase or decrease growth depending on the types of plants involved and the resources on the landscape. These interactions also mean that distance is best treated \textit{implicitly} in modeling the effects of the interactions, rather than \textit{explicitly} as it is in most location problems. Both heterogeneity and locality generate feedback conditions, whereby the positioning of a plant on a landscape changes the resources available at that location. The placement of a tree, for example, changes the light available nearby, which then affects the growth of any plants located in those modified conditions.

The model presented here captures the domain-specific features of this problem using an agent-based strategy. In this model, each plant is an agent that ``lives'' on a simulated landscape. Agents have light and water requirements for growth, selected to match the requirements of real plants, and a growth curve that determines agent fitness in simulated light and water conditions. A multiobjective fitness function captures the trade-off between maximizing plant agent growth on the landscape and minimizing each agent's water use. The total water needed on the landscape is the emergent property of this individual behavior.

Two optimization strategies commonly used on location problems --- simulated annealing and a genetic algorithm --- are applied to find the locations on the landscape that maximize the fitness score. These strategies are compared to an agent-based optimization routine that is designed to capitalize on the novel features of this domain. In this approach, agents maximize their own individual fitness instead of the global fitness on the landscape. This distributed strategy produces solutions comparable to the genetic algorithm, but in a fraction of the time required for that algorithm.

his work also includes experiments with live plants to generate the data for the agent growth curves and to validate the performance of the optimized arrangements. The growth curves were generated using two species with different light and water requirements grown under a range of light and water conditions. A validation experiment, where live plants were arranged in a random and an optimized configuration, shows the value of this optimization approach for water conservation. Both arrangements had zero mortality and all plants looked healthy throughout the experiment. However, the optimized arrangement used significantly less water.