Source: NORTH CAROLINA STATE UNIV submitted to NRP
DSFAS-AI: AGRICULTURAL DECISION INTELLIGENCE MODELING SYSTEM FOR HUMAN-AI COLLABORATIVE ACTION ELICITATION AND IMPROVEMENT (DECIDE-SMARTER)
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
ACTIVE
Funding Source
Reporting Frequency
Annual
Accession No.
1028503
Grant No.
2022-67021-37137
Cumulative Award Amt.
$649,722.00
Proposal No.
2021-11509
Multistate No.
(N/A)
Project Start Date
Jun 15, 2022
Project End Date
Jun 14, 2026
Grant Year
2022
Program Code
[A1541]- Food and Agriculture Cyberinformatics and Tools
Recipient Organization
NORTH CAROLINA STATE UNIV
(N/A)
RALEIGH,NC 27695
Performing Department
Computer Sciences
Non Technical Summary
Artificial Intelligence (AI) advances have revolutionized industries (e.g., manufacturing, medicine); yet, the power of AI remains elusive for most agriculture stakeholders. This is exacerbated by technologists, which tend to focus on technology and solutions, as opposed to stakeholders and problems. This project will advance Decision Intelligence (DI) as a stakeholder- and problem-first approach. DI, which has been used effectively in industry for decades but not formalized academically, will enable new ways to link AI technology solutions with stakeholder expertise and domain-specific models, visualization tools, and built-in methods of identifying biases. Through participatory design with stakeholders, and using the sweetpotato supply chain as a use-case, the project's objectives are to: 1) create software for visualizing data-driven decisions; 2) conduct experiments to improve user decision making; 3) identify sources of bias that would reduce technology access and adoption; and 4) develop an open-source software reference architecture for DI tools. This project will provide a formalized methodology for users to identify and implement technology solutions in evolutionary and tractable ways, and addresses the DSFAS-AI topics of facilitating real-time decision making, incorporating new methods to reduce bias in machine learning methods, and developing open-source platforms for improved adoption of AI tools. With increasing innovations in on-farm AI capabilities, a rising tide of data is primed to overwhelm stakeholders without proper tools. Stemming the tide will require innovation in sensing, AI, software, cognition, and agriculture--all areas the project team comprises expertise--indicating the project is timely and the team is well-suited to meet producers' needs.
Animal Health Component
0%
Research Effort Categories
Basic
50%
Applied
0%
Developmental
50%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
2051450208030%
4027410208070%
Goals / Objectives
The DECision Intelligence moDEling System for huMan-AI collaboRative acTion Elicitation and impRovement (DECIDE-SMARTER) in Agriculture project is designed to lay the foundations of democratized access to Decision Intelligence (DI) technology for stakeholders across the agri- culture value chain. Nearly 10 decades of developments in Artificial Intelligence (AI) have yielded incredible breakthroughs that have revolutionized industries from manufacturing, to medicine, to automotive, and beyond; however, the power of those advancements remains elusive for most, re- quiring highly-specialized technical skills, access to data and computing resources, and software that is out of reach for most stakeholders. Further, much of the progress in AI has followed a similar path: the focus has been on developing highly-specialized solutions to relatively simple decisions in isolation. Real-world decision making is rarely so simple. Especially in agriculture value chains, multiple stakeholders with individual goals and circumstances are making many de- cisions with implications in both space and time (e.g., which fields to plant with which varieties of seed and when) and with heavy influences from external factors (e.g., market pricing, climate change, weather events).DI is a practice of understanding, modeling, simulating, and improving decision making in complex, multistakeholder environments that has decades of demonstrated success in a variety of industries from telecommunications, to finance, to public health; however, like AI, DI remains elu- sive to most stakeholders as the expertise required to implement DI methods is highly-specialized and, as a general rule, only available to customers or businesses with deep pockets that can con- tract for customized solutions. The DECIDE-SMARTER in Agriculture project will change this landscape by providing the software support that will make DI available to a broad range of stake- holders without the need for technical expertise, expensive and complicated computing infrastruc- ture, or even extensive sources of data. Through a process of participatory design, the DECIDE- SMARTER project team will work with stakeholders in the sweetpotato value chain to: 1) char- acterize, understand, and leverage human cognitive abilities in the design of software interfaces to visualize decisions; 2) identify potential sources of bias in the DI process that would present barriers to democratized access to the technology; and 3) develop a reference architecture and pro- totype implementation of a modeling, simulation, and visualization framework for implementing DI models with agriculture stakeholders. To facilitate these contributions, the DECIDE-SMARTER team will leverage the ongoing research, data acquisition, and stakeholder efforts by the Sweet- potato Analytics for Produce Provenance and Scanning (Sweet-APPS) team, a multi-disciplinary endeavor that aims to reduce agricultural waste and maximize yield for North Carolina's sweet potato growers.
Project Methods
There are three philosophies guiding the approach taken in the DECIDE-SMARTER project:Stakeholder interaction is of paramount importance through all phases of research and software development to ensure their needs are addressed.Iteration is critical to ensure that the ultimate products delivered truly meet the needs of stakeholders across a wide variety of communities.Evaluation of all project activities at regular points in the project lifecycle will provide valuable insight into progress and both long- and short-term future directions.The first six months of the project will emphasize planning and conducting stakeholder interaction workshops to obtain a preliminary set of requirements for the DECIDE-SMARTER software and initial Causal Decision Models. In parallel with workshop planning, developing a "discovery prototype" of the DECIDE-SMARTER software built from "off-the-shelf" components selected to support use cases identified through stakeholder discussions will commence. We will conduct the workshops with existing stakeholder groups from NC Sweetpotato Field Days, organized by NCSU extension faculty affiliated with the Plant Sciences Institute.The DECIDE-SMARTER project will be organized into six cycles of innovation, each with a regular schedule of planning, development, and evaluation that will be punctuated with stakeholder interactions at strategic points in the process. This iterative approach to managing the project will follow the well-established Agile methodology which is considered to be best practice in software development. In addition to the agile development of the DECIDE-SMARTER software architecture and prototypes, we will also use Agile methodology for overall project management.Building on the discoveries from each iteration through the continuous improvement cycle, we will introspect, diagnose, and plan improvements to enhance all aspects of the project's activities, from the composition of stakeholder groups and types of interactions to fundamental research goals, processes, and outcomes. Using a combination of internal evaluations by project personnel (i.e., surveys and stakeholder advisors), we will obtain diverse feedback. Where appropriate, quantitative measures of performance will be used to adjust processes, goals, and methods. These evaluations will also be used as part of regular internal progress reviews and as part of the process for prioritizing and sunsetting activities. The six cycles of innovation in this project will be organized into planned "sprints" as follows: 1) post hoc analysis of prior cycle (0.75 months), 2) planning for new goals and deliverables (0.75 months); 3) iterative development, integration, internal testing, and internal evaluation (four 1-month sprints); and 4) deployment and external evaluation (0.5 months). The six-month iterations coincide with stakeholder engagement activities, including sweetpotato field days, which maximizes opportunities for the DECIDE-SMARTER team to receive stakeholder feedback.The project objectives include 1) prototyping visualizations, 2) developing a reference software architecture, 3) integrating research advances and existing AI assets into an implementation of the reference architecture, and 4) developing a framework for identifying biases in DI processes and tools. Evaluating these objectives will occur primarily through quantitative and qualitative user studies, focus groups, and stakeholder interviews.

Progress 06/15/23 to 06/14/24

Outputs
Target Audience:Out efforts during the reporting period have emphasized three target audiences: Sweetpotato producers, specifically packers. We worked with packers at our partner institution Scott Farms to develop a Causal Decision Model for packing decisions. Technical people interested in mathematical, statistical, and computational forecasting. More specifically, readers of the journal Foresight. Practitioners of Decision Intelligence, including a wide range of students, faculty, business executives, software engineers, and analysts. Changes/Problems:The most significant change to the project is that study participants have universally preferred to schedule teleconference, rather than meet in person. The funds we had budgeted to support travel for in person studies have been requested to be reallocated to enhance other aspects of the project. Because we've been able to conduct the studies remotely, there has been (nor do we anticipate there to be) any impact to the project scope and deliverables. What opportunities for training and professional development has the project provided?For the project reporting period, two graduate students have had professional development opportunities. Both students have benefited from having regular and sustained interactions with a sizable team of senior personnel across three institutions. They have also designed and run user studies to gain a formative understanding of the cognitive underpinnings of Decision Intelligence methods. These efforts have led to one MS Thesis. How have the results been disseminated to communities of interest?Results have been disseminated in three ways: 1) a peer reviewed journal article in Foresight: The Journal of Applied Forecasting; 2) the launch of the opendi.org website, which contains resources for people learning of Decision Intelligence technologies and serves as an access point to the opensource architecture and reference implementations; and 3) through travel to a workshop and the associated discussions about this project. What do you plan to do during the next reporting period to accomplish the goals?Our main efforts during the final reporting period will be to mature the reference architecture and reference implementations; develop and implement a library of "controls" for converting a Causal Decision Diagram to a runnable Causal Decision Model; run additional user studies on the completed, runnable, sweetpoato packing model; and to further disseminate our results in additional publications and through OpenDI initiative. As a related goal, we aim to grow the OpenDI initiative to attract additional contributors that will make tools more useable and useful for key agriculture stakeholders.

Impacts
What was accomplished under these goals? During the project reporting period we have emphasized two of the overall project goals: 1) eliciting and developing a decision model for sweetpotato packers, and 2) releasing an opnsource reference architecture standard (and building an associated community of contributors). For the first goal, we have worked with partners at Scott Farms in North Carolina to develop a Causal Decision Diagram governing their daily packing decisions. We've gained an understanding of the nuances of the challenges they face, as well as the technology they currently use (spreadsheets). The Diagram has proven a useful tool for reaching cognitive convergence on the outcomes they desire, the actions available to them, as well as the external and intermediate factors that influence and mitigate the effects of their chosen actions on outcomes. This model has become the foundation upon which we've begun studying visualization paradigms (which will be a bigger emphasis for the next project reporting period). For the second goal, we have launched the Open Decision Intelligence Initiative (OpenDI). OpenDI is conceived to be the governing organization for creating standards that promote interoperability between tools developed for Decision Intelligence. This standard will lower the barrier to entry, drive adoption, and help to democratize access to DI technology. The OpenDI initiative, lead by PD Roberts and Co-PD Pratt (as well as one external collaborator) has created a website to curate DI concepts and disseminate tools, techniques, standards, and knowledge. As of the reports due date, the website (and the initiative more broadly) is just getting off the ground. However, it's social media presence on Discord has already generated more than 30 individuals to sign up.

Publications

  • Type: Other Journal Articles Status: Accepted Year Published: 2024 Citation: Pratt, Lorien, David Roberts, Nadine Malcolm, Brian Fisher, Katie Barnhill, Daniela Jones, and Michael Kudenov. How Decision Intelligence Integrates Forecasting, AI, and Data into Complex Decisions. Foresight: The International Journal of Applied Forecasting, no. 72 (January 1, 2024): 5257.


Progress 06/15/22 to 06/14/23

Outputs
Target Audience:The DECIDE SMARTER project has conducted activities with stakeholders in the North Carolina Sweetpotato supply chain. Our efforts have emphasized interactions with three populations: 1) research faculty, 2) extension faculty, and 3) growers and packers. We have interacted with these groups both in person and virtually following two different IRB approved protocols. Our initial efforts have included four Causal Decision Diagram elicitation sessions---one with research faculty, two with extension faculty, and one with a combination of growers and packers. We also pilot tasted a prototype runnable decision model at the NC State extension Sweetpotato Field Day where industry representatives had an opportunity to discuss the project and interact with a prototype Causal Decision Model. Changes/Problems:We have made more substantial and rapid progress on decision elicatations and model building than we intially anticipated, and thus are currently working on a rebudet to more rapidly move forwared software engineering efforts. This doesn't represent a change in priorities, but a change in the timing of the priorities. What opportunities for training and professional development has the project provided? Nothing Reported How have the results been disseminated to communities of interest?The main dissemination activity during the first seven months of the project has been via the stakeholder interactions that occurred during the four decision elicitations as well as the interactions with producers at the Sweetpotato Field Day. What do you plan to do during the next reporting period to accomplish the goals?Because of the successful decision elication sessions, we have acquired the data needed to move forward with implementation of additional protype runnable models and to begin cognitive studies of decision visualization technologies. The next 4-6 months will emphasize dissemination via two publications we have in preparation as well as technology development. The first publication will address the decision model we created for management of nematode outbreaks in sweetpotato crops. The decision intelligence process proved to substantially improve understanding among those who particiapted, but also critically highlighted gaps in scientific knowledge about nematode management effectiveness. The second paper in preparation will address initial effort to survey literature in bias in technology and Artificial Intellgience, and highlight unique challenges in Decision Intelligence that aren't easy addressed by methods developed for Artificial Intelligence. Technology development efforts over the next reporting period will emphasize prototype visualizations for decision models using portable javascript technologies that will support future development and integration into a web application for decision modeling. The aim is to produce 10 "decision control" visualizations that will support the implemtnation of two runnable models based on two of the four decision eliciation outputs. Additionally, we will complete work on the design of the DECIDE SMARTER reference architecture and begin implementaiton of the backend APIs. The later half of the reporting period will emphasize running additional decision elictation sessions and evaluating new runnable models implemented using the decision controls we develop and informed by the initial results of the cognitive studies.

Impacts
What was accomplished under these goals? The first seven months of the DECIDE SMARTER project have emphasized relationship building and data collection to support future cognitive studies and technology development. Our main progress has come in the form of completing four separate decision elicitation sessions across three different stakeholder groups as well as running an evaluation of a prototype model at the NC State Sweetpotato Field Day. These decision elicitation sessions have yielded the first known data sets in existence that will support the next phase of technology development, bias identification, and cognitive studies. Additionally, reducing one of the elicited decision diagrams to a runnable model that we presented for initial feedback from stakeholders at the Sweetpoato Field Day has also given us valuable information about producers' attitudes towards technology, and critically the type of decisions being made. Lastly, we have made progress on documenting the reference architecture for the Open Decision Intellgience initiative, for which the DECIDE SMARTER software will be the foundation. The design document draft covers both user experiences as well as technical artifacts that will comprise the implemented software system.

Publications