The Agricultural DECision Intelligence moDEling System for huMan-AI collaboRative acTion Elicitation and impRovement (DECIDE-SMARTER) project will lay the foundations of democratized access to Decision Intelligence (DI) technology for stakeholders across the agriculture value chain, filling a longstanding gap between technology and decision makers. Through a process of participatory design, the project team will work with stakeholders in the sweetpotato value chain to: 1) Create a software asset that helps growers with an otherwise difficult decision; 2) conduct experiments that inform the best software interfaces possible to support complex agricultural decision making (through characterizing, understanding, and leveraging human cognitive abilities; 3) identify potential sources of bias in the DI process that would present barriers to democratized access to the technology; and 4) develop a reference architecture and prototype implementation of a modeling, simulation, and visualization framework for implementing multiple DI models with agriculture stakeholders. The project will leverage the ongoing research, data acquisition, and stakeholder efforts by the Sweetpotato 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.

The purpose of this agreement is to utilize our existing African swine fever (ASF) transmission model to create a website dashboard that calculates the number of certified swine sample collectors (CSSCs) needed to effectively respond to an ASF outbreak in the U.S. Intended beneficiaries include the creators of a national CSSC training program, the National Pork Board, as well as state and federal animal health officials and industry veterinarians, who currently lack information about the number of CSSCs needed for surveillance and movement permitting within outbreak control zones. This is particularly challenging due to differences between ASF strains, which lead to epidemics of varying duration and magnitude and could be catastrophic if not effectively controlled. Activities to be performed will build on the success of previously funded NADPRP projects that have supported the development of a transmission model (PigSpread) for low, moderate, and high virulence ASF strains, and the collection of vast movement datasets from swine businesses across North Carolina, South Carolina, and Virginia. Specifically, we will modify our existing ASF model to calculate the number of CSSCs needed based on factors such as movement networks, time taken to collect and deliver samples to surveillance laboratories within an 8-hour workday, and mandatory downtime between farm visits among other biosecurity rules. We will subsequently construct a dashboard that swine industry stakeholders can use to input information to generate region-specific information for their CSSC training programs. This project will deliver a website dashboard that calculates the number of people that need to be trained via state CSSC training programs based on key transmission factors such as ASF strain incubation and latency, as well as regional-specific information such as farm and laboratory locations and movement networks. In doing so, this project will help states and swine companies conduct strategic CSSC training to ensure sufficient blood samples can be collected from swine farms during an ASF outbreak, thus enabling effective disease surveillance and distribution of movement permits within outbreak control zones. Ultimately, this has the potential to help preserve swine business continuity in the event of an ASF emergency in the U.S.

Investigators at Oak Ridge National Lab (ORNL) are interested in measuring physiological characteristics of dogs as they are exposed to environmental stimuli. We will develop 1) custom wearable sensing units capable of aggregating measurements from working dogs in controlled environments during experiments, 2) corpora of data from experiments conducted by ORNL in a format amenable for statistical analysis from ORNL, and 3) software to enable ORNL to extract and process the data describing those physiological measurements automatically. We will provide input into experimental design to facilitate effective use of sensing devices for analyzing physiological responses.

We propose to develop tools and techniques that will enable more effective two-way communication between dogs and handlers. We will work to create non-invasive physiological and inertial measuring devices that will transmit real-time information wirelessly to a computer. We will also develop technologies that will enable the computer to train desired behaviors using positive reinforcement without the direct input from humans. We will work to validate our approach using laboratory animals in the CVM as well as with a local assistance dog training organization working as a consultant.

Machines learning from feedback is a well-studied problem. If robots and software systems can successfully adapt, they can remain useful in changing environments, in situations unanticipated at design time, and can take direction from human users. This proposal contributes a new algorithm, Income Learning (I-learning), that is designed to thrive in these scenarios. The emphasis of the work will be on theoretical and empirical analyses of how I-learning and existing temporal difference methods (that maximize the expected reward) differ in performance on a variety of tasks, and how I-learning is better able to take advantage of human teaching.

DO6 Sensemaking

DO6 Sensemaking

We propose to develop techniques that will enable humans to train computers efficiently and intuitively. In this proposed work, we draw inspiration from the ways that human trainers teach dogs complex behaviors to develop novel machine learning paradigms that will enable intelligent agents to learn from human trainers quickly, and in a way that humans can intuitively take advantage of. This research aims to return to the basics of programming---it seeks to develop novel methods that allow humans to tell computers what to do. More specifically, this research will develop learning techniques that explicitly model and leverage the implicit communication channel that humans use while training, a process akin to interpreting the pragmatic implicature of a natural language communication. We will develop algorithms that view the training process as an intentional communicative act, and can vastly outperform standard reward-seeking algorithms in terms of the speed and accuracy with which human trainers can generate desired behavior.

LAS DO5 Task 5.3 Sensemaking

DO 2 Task 3.4 Activities