An artificial language for data-driven self-adaptation of networked robots in dynamic environments

The interactive dynamics of goal-oriented multi-agent networked robots with on-board sensing, computation, and actuation devices, present a complex distributed computational environment of high dimensionality. The generating physics of such a system operating in an uncertain environment can be adequately captured in an artificial language that expresses the causal patterns observable in sensor data with maximal compression while preserving the statistical predictability of system states under Markovian assumptions. Hence it enables time-constrained in-situ distributed computation, communication, and data-driven adaptive control in resource-constrained uncertain operational environments. The multivariate sensor data is partitioned and symbolized for deriving the alphabet of the language. Observed data from multiple sensors is expressed as a univariate sequence of symbols from this alphabet. The semantics of the language are extracted from the observed data streams as invariant patterns which capture the essential causal structure of the dynamic system. An undersea mine-hunting mission using an undersea robot with on-board side-scan sonar is used to illustrate the development and use of this physics-driven computational language for time-constrained situational awareness and adaptive control.