Overview

This project explores the application of Partially Observable Markov Decision Process (POMDP) models to formulate optimal conservation strategies for Florida Manatee populations. The work demonstrates how decision-making frameworks can address wildlife conservation challenges under uncertainty, providing a simulation-based approach to optimize resource allocation for conservation efforts.

Problem Context

Florida Manatees are a threatened species facing multiple environmental challenges including:

• Climate change effects on marine ecosystems
• Loss of seagrass food sources
• Unusually cold winters
• Algal bloom outbreaks

The project addresses the resource management challenge of determining optimal conservation strategies under uncertainty, where the exact population state is difficult to observe directly.

Technical Approach

POMDP Framework

Partially Observable Markov Decision Processes (POMDPs) are mathematical frameworks for sequential decision-making under uncertainty where:

• The true state of the system is not directly observable
• An agent receives noisy or incomplete observations
• Actions affect future states probabilistically
• The goal is to find an optimal policy that maximizes long-term rewards

This framework is particularly well-suited for conservation problems where population states are uncertain and management decisions must be made based on limited observational data.

Problem Formulation

Our simulation model defines:

• States: [extinct, endangered, threatened]
• Actions: [do nothing, survey, rescue, rehabilitate]
• Observations: [see nothing, wild carcass, wild living]
• Initial State: threatened
• Discount Factor: 0.9

Note: This is a simulation based on available literature and our own problem formulation. Transition probabilities, observation models, and reward structures were designed based on educated assumptions about manatee conservation dynamics.

Solution Methods

Offline Methods:

• QMDP: Simple approximation assuming perfect observability
• Fast Informed Bound (FIB): Tighter upper bound incorporating observations

Online Methods:

• PO-UCT: Partially Observable Upper Confidence Tree combining Monte Carlo Tree Search with Upper Confidence Trees

Key Results

• Generated alpha vector visualizations showing optimal action regions based on belief states
• Created ternary plots mapping full belief states to optimal actions
• Demonstrated that rehabilitation is optimal for moderate endangerment probabilities (0.071-0.500 for QMDP)
• Survey actions dominate when endangerment probability is low
• Rescue actions are preferred when endangerment probability is high
• POMDP methods can help as a framework for wildlife conservation decision-making

Technologies Used

• Julia programming language with Pluto notebooks for interactive development
• POMDPs.jl ecosystem including:
• QuickPOMDPs.jl
• QMDP.jl
• FIB.jl
• BasicPOMCP.jl
• D3Trees.jl for visualization

Paper Link

Github Link

Project completed for CS 238 (Decision Making under Uncertainty) at Stanford University. Collaborated with Rubens Lacouture and Aarya Mecwan.