Experimenting Systems: An Experiment as a Logic Unit to Build an Exploration Logic

Val Bykovsky, Space Dynamics Lab, Utah State University, Logan, UT, val.bykovsky@gmail.com

Direct experimentation was always a primary source of knowledge by providing the data dependencies between the conditions of an experiment and its results. When properly generalized by human expert, the dependencies allow for building a mapping between the physical variables, often in the form of an equation, a well-known prediction mechanism. However, the performance and logical analysis depth are quite limited due to the presence of a (slow) human element in the knowledge-acquisition loop. So, we are looking into direct computer-driven methods of high-performance real-world knowledge acquisition. We discuss the Experimenting Systems (ES) designed for exploration using automated measurements of dependencies. The focus is on "experimenting" vs. "computing" systems. What makes difference is that the ES has configurable sensors controlled by the exploration program. In a computing system, the input "data" is the only link with real world, and all the real-world dependencies are being simulated computationally. An ES directly interfaces the real world, and the necessary dependencies are measured directly and on-demand if need. We propose to use the dependencies as training units to build an input-output mapping directly in computer using program-driven generalization and a generic framework to be structured by training data. Such a mapping becomes a direct, dependency-based prediction mechanism. No equation is required. Direct experimentation is complexity-neutral - it can deal with any objects no matter how complex. We propose and discuss the computer-driven environment for high-performance experimentation. It has its roots in "experimental mathematics", a term coined by Nick Metropolis in late 40s, and in data-driven modeling also proposed by Nick Metropolis who was the first to implement a two-way connection between a computer (at the University of Chicago Institute for Computer Research, late 60s) and the Navy cyclotron. It also has interesting connections to the Monte-Carlo method invented jointly by S. Ulam and N. Metropolis (in late 40s) and extends the original physics-guess-based sampling to the demand-driven real-world online experiments. The proposed ES allows one to build programmatically trial objects, setup the sampling options, run the experiments and measure the properties of interest. Basically, this is a search in physical configuration space for common features. With modern computers and high-speed multi-facet sensors, search can be effective and fast. The experimentation approach is a kind of "calculus of dependencies" and has little to do with computational approach to problem-solving. The experimentation environment, its features and examples of experimentation are briefly discussed.