Santa Fe Institute James P Crutchfield P I

Santa Fe Institute: James P. Crutchfield, P. I. Dynamics of Learning & Distributed Adaptation CAHDE REF ACFC: Adapting to instabilities in air flow control Air. Ops: Emergence of spontaneous leadership Solution: Interacting reinforcement and -machine learning agents solve a group task Approach: Pattern Discovery: Beyond pattern recognition Design & analysis based on sound principles of learning Metrics for cooperation in large-scale systems Multi-Agent System Science (MASS) Design Approach: Learning Dynamics for the Emergence of Adaptation in Distributed Systems Critical Elements: Coordination, adaptation, uncertainty Problem: Analyze distributed information and coordination mechanisms in large-scale MASs Metrics: Stored information, causal synchrony-distributed information, predictability, generalization ability CAHDE Results Predictive theory of agent learning: Agent Complexity v. Prediction Error v. Data Size Pattern Discovery: The “Aha” Effect Incremental learning algorithm How env’t structure leads to unpredictability for agent Synchronization for chaotic environments and agents: Predict required data and time to synchronize Transient information: New metric of synchronization Causal Synchrony: Detect coherent subgroup behavior Dynamics of reinforcement-learning agents: Nash equilibria v. oscillation v. chaos Demonstrations Causal Synchrony for large-scale MASs: Biologically realistic neural network Coordination Game MAS (Minority Game, USC) Toolkit v 1. 0 Causal-State Splitting Reconstruction Algorithm Library for Estimating MASS Metrics Game-Theoretic Simulation Platform for RL MASs Future Plans Solvable MAS systems : Continuous-state, -time agents Dynamical theory of how learning and adaptation occur Monitor emergence of cooperation in agent collectives Test on in-house autonomous robotic vehicle collectives