One of the most important unsolved problems in artificial intelligence (AI) is how to plan for the future given incomplete information about the present. Doing so requires an ability to adapt quickly to changing circumstances beyond the range of eventualities identified in past training data. A novel Lawrence Livermore National Lab (LLNL) computational method is critical for the Department of Defense when they deploy AI-enhanced capabilities in real-world operations. It’s in these operations that new variables are created at an exponential rate in real-time, making it difficult for traditional AI to keep pace.
Another challenge faced by AI is how to synthesize and contextualize data as it’s acquired and then taking action in new, unforseen scenarios. Current Large Language Model (LLM) approaches cannot do this because they require vast training data and expensive offline computing resources.
A recent LLNL machine-learning breakthrough identified new ways to solve reinforcement-learning problems by casting them as components of completely integrable dynamical systems. This approach enables efficient solutions through parallel computations.
Additional advances in probabilistic machine learning and high-throughput computing methods also provide decision-support solutions for problems under unique national security mission constraints. Decision superiority researchers at LLNL — including data scientists, statisticians, physicists and mathematicians — work with the Global Security Directorate to apply these innovations by partnering with experts across the Lab and other national laboratories in defense-systems modeling and simulations.
