Bio-synthetic lifeforms: a framework for generating evolvable, digital metacognitive agents and for using their structure as a blueprint to produce novel synthetic lifeforms.
Abstract
While AI//ML systems can learn from data and adapt their behavior within the predefined problem space in increasingly sophisticatedmanner, they still do not possess the capability to autonomously evolve beyond those boundaries. They are constrained by human-imposed design biases, the choice of algorithms, comprehension limitations and training data limitations. Our aim is to develop a framework for generating evolvable agents that can create and evolve their own concepts, combine them into complex conceptual networks, and self-analyze and assess their execution. With that, agents within the framework will be able to continuously learn how to conceptualize and understand the environment, create subjective model of the world and use that model to plan and assess their own actions. Through experience they will gradually improve their functioning and adapt to changes by developing new concepts and expanding theirknowledge base, using diverse sources they are exposed to. Therefore, they will be able to make decision in dynamic environments that have no clearly defined rules.These software systems can be kept as purely digital agents, but our long-term vision is to developprocedures for using their evolved structure as blueprint to produce novel synthetic lifeforms.In this project we will develop a preliminary mathematical framework, implement and demonstrate a preliminary software proof-of-concept, develop a roadmap for further development and identify use cases. Evolvable software agents will be able to build concepts using minimal available data and will have implemented architecture for association of elementary concepts into higher-order concepts and stable conceptual networks.In the first set of prototypes, we will focus on pattern recognition so the agents can be specialized for automated visual scene understanding. Therefore, the platform can be used for developing agents for Autonomous agent perception, and for Understanding surveillance imagery, that are outlined as one of the "Research Concentration Areas" within C5ISRT s Division 311 MCIS "Machine learning, reasoning and intelligence" thematic area. Further development steps will bring our platform to the level of generating self-regulated agents with the ability for enhanced understanding, analysis and generation of new viewpoints about observed complex systems. It will make it suitable for the Command and Control program, especially for the "Automated Information Integration Across Enterprise Sensors and Information Stores" research concentration area.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 08, 2024
- Source ID
- N629092412024
Entities
People
- Igor Balaz
Organizations
- Office of Naval Research
- United States Navy