Topic

OUSD (R&E) MODERNIZATION PRIORITY: Cybersecurity; Network Command, Control and Communications; Artificial Intelligence/Machine Learning; General Warfighting Requirements TECHNOLOGY AREA(S): Weapons; Information Systems; Battlespace OBJECTIVE: Creation of a Modeling and Simulation (M&S) development and execution environment which significantly decreases the time to execute statistically significant batches of stochastic simulation runs for the purpose of estimating scenario output/outcome distributions while improving our knowledge of the outcome distributions. DESCRIPTION: This topic seeks the development of a paratemporal M&S architecture that must not only execute efficiently, but it must be scalable and should not introduce any significant burden on developers of the simulations and the underlying algorithmic models. Commonly, to estimate the distribution of scenario outputs/outcomes of a stochastic simulation, we would execute the simulation repeatedly with different random seeds until there were sufficient runs to estimate the distributions of the results to an acceptable level of confidence. In “cloning”, “5-D” or “paratemporal” simulation, a simulation is run until a stochastic decision point is reached, at which point the current simulation states and probabilities of branching are saved. The simulation is then executed for one possible branching until another branching point is reached. The process is then repeated until the execution end-point is reached. The simulation is then restarted at a saved state (cloned) and an alternative branch is taken. In this way, the simulation need not be completely rerun for additional executions and the probabilities and distributions of each path execution can be calculated. This is also extremely parallelizable for great efficiency in execution on multiple processors. For an academic “toy” example, assume a simulation that starts at execution time t=0 at state s=0. After one second of computational time (t=1) a random model is reached which either adds 1 to the state, or 0 resulting in states of either s=1, or s=0. This repeats until instantaneously after t=4, at which point there are a total of 16 possible branching routes with the final states being (4, 3, 3, 2, 3, 2, 2, 1, 3, 2, 2, 1, 2, 1, 1, 0), with probabilities calculated with stored probabilities at each branch. Starting from the beginning but intentionally choosing outcomes to step through each possible branching route would require 64 seconds of computational time and gives complete knowledge of the distribution of results. In this “toy” example, a statistically significant number of executions (>20 runs/>80 s of execution time) actually requires more resources while providing less knowledge of the output distribution. Using “paratemporal” simulation that reuses results up to new branch points, all 16 possible branching routes can be executed in 14 s. A more sophisticated paratemporal execution that recognizes and reuses tree branches from common states could reduce this execution time further to 7 s. However, simple paratemporal simulation suffers from a scalability issue. Paratemporal simulations with a small number of branchings (e.g. missile intercept draws) and necessitated segment state saves have been demonstrated. However, a simple implementation of paratemporal simulation does not scale as the number of branches increases (e.g. numerous detection vs. noise draws in sensor simulations). To tackle this scalability issue, it is desirable to develop more sophisticated paratemporal simulation techniques that take advantage of uncertainty quantification methods to increase the scalability while still providing improved, albeit imperfect, knowledge of scenario outcome/output distributions. For example, it may be possible to combine segments of similar draws (such as the sensor detections) into super segments to minimize state saves and select branch/branch family executions to maximize knowledge of outcome distributions. Any solution will almost certainly require a communication from an algorithmic model using a random draw to the simulation engine/execution manager to inform it of the branching point, provide the branch probabilities, and store state information. It is critical that the developed solution minimize the complexity of developing the M&S software implementing this model to the simulation engine interface. The developed architecture must not only execute efficiently, but it must be scalable and should not introduce any significant burden on developers of the simulations and the underlying algorithmic models. PHASE I: Develop a proof-of-concept of the algorithms, tools, software and analyses that demonstrate potential for achieving the topic objectives: - Stochastic batch speed-up (via speed test comparisons vs. traditional) - Increased knowledge and confidence in scenario outcome/output distributions (via test & analysis comparing resulting distributions and confidence intervals with traditional methods, including marginal gain with more executions) - Scalability as number of random draws increases (via demonstration & analysis with increasingly complex stochastic models/simulation/scenarios) - Minimized development and maintenance workload for models and simulation infrastructure using the developed capability (via demonstration and analysis of required software tasks) PHASE II: Develop a full prototype capability demonstrating initial capabilities per topic objectives with the intent in testing the capability for experimentation in government M&S labs. This should include prototype level user and design documentation. Development should facilitate cyber security approval for loading the prototype software on government computer systems through cyber aware design decisions and development of cyber security artifacts. PHASE III DUAL USE APPLICATIONS: Develop operational capability for use in government simulations, including user and design documentation. Maintain and improve capabilities based on Phase II and Phase III use experience. Continue to support cyber assurance. REFERENCES: 1) S. B. Yoginath, M. Alam and K. S. Perumalla, "Energy Conservation Through Cloned Execution Of Simulations," 2019 Winter Simulation Conference (WSC), 2019, pp. 2572-2582, https://ieeexplore.ieee.org/document/9004821 2) Xiaosong Li, Wentong Cai, and Stephen J. Turner. 2017. Cloning Agent-Based Simulation. ACM Trans. Model. Comput. Simul. 27, 2, Article 15 (July 2017), 24 pages. DOI: https://doi.org/10.1145/3013529 3) C. Lammers, J. Steinman, M. Valinski, K. Roth. “Five-Dimensional Simulation for Advanced Decision Making,” SPIE—Enabling Technologies for Simulation Science XIII, Paper SPIE 7348-16. KEYWORDS: Models; Simulation; M&S; Simulation Engines; Stochastic Models; Random Models; Uncertainty; Probability; Statistics; Simulation Cloning; Cloned Execution; 5-D Simulation; Five-Dimensional Simulation; Paratemporal Simulation; Simulation States