Prev-Talk | Next-Talk | All-Talks | Talks-Sorted | MOOS-DAWG'24
Talk-12: Siren’s Song: Strongly Typed Composable Ontologies with Dynamically Re-Trainable Classifiers for Addressing Knowledge Acquisition Challenges for Autonomous Systems
Erik Thomsen, CSO Sensepoint, Inc., Cambridge MA
Self-navigation, e.g., how to steer, how to maintain a course, and how to avoid or navigate around obstacles, has, until recently, been the focus of most work on maritime automation. But that is only a part of the automation story. Boats are self-navigating for a purpose; the most common being ‘embodied’ knowledge acquisition e.g., shadowing or identifying a target, preparing for the environment, or discovering patterns of life.
The challenges of embodied knowledge acquisition include converting sense data and telemetry into hypotheses about the world, assessing the warranted confidence in current sense data, deciding what observation(s) would provide the greatest mission information value, and generating plans to improve that value.
Standard techniques for representing knowledge (regardless of how sourced) are based on ontologies. The problem with standard ontologies is they are weakly typed, limited to representing schemas, and limited to Boolean ANDs as a means of composition. We use a new kind of ontology that is conformant with the DoD standard BFO and strongly typed with a compositional grammar capable of assessing and driving ML processes.
In this presentation, using UsV target identification as an illustrative test case, we show how a strongly typed composable ontology can structure all parts of the embodied knowledge acquisition process and enable an autonomous vehicle to make intelligent navigational, sensor control, ML retraining and communications decisions in order to acquire relevant knowledge.
Categories:
- Multi-Agent Decision-Making
- Adaptive Control