Agent Based Learning Systems Interfacing and Validating Models of the US Army TRAC Tactical War Game Deborah Duong Christopher Bladon Agent Based Learning Systems AHFE 2012
SIMmiddleware Translates Disparate Data between IW Models SIMmiddleware enables analysts to apply IW data from one unique situation in the world to a different but unique situation in a model. Input data, data traded between federated models, calibration data, and testing set data all have uncertainty of match as well as translation difficulties because of the different concepts used to arrange different data. SIMmiddleware translates data to ensure one model “means the same thing” as another model. If there is no exact match, SIM translates data probabilistically. SIMmiddleware enables hybrid modeling. Analysts can couple models “loosely” , at the level of general patterns, or tightly, at the level of details. SIMmiddleware can integrate models at different levels of aggregation, for example, “tactical” and “operational/strategic”. SIM enables data to be compared “apples to apples”. Validation needs to be at the level of statistical patterns rather than single outcomes. The real world is just one possible world, and simulations model many possible worlds. We know a simulation is good if what is rare in the simulation is rare in the world, and what is common in the simulation is common in the world (under right circumstances). SIMmiddleware enables analysts to express data in statistical patterns and dynamics. After SIMmiddleware translates data to a common lexicon, SIMmiddleware compares data at the level of statistical patterns and dynamics to calculate a ‘distance’ score that measures the statistical distance between dynamic patterns. SIMmiddleware’s distance score allows comparison of different versions of the same model, different scenarios, and models against data for a calibration or a validation score. SIMmiddleware can validate a model against multiple real-world data sources with uncertain matches to the model. Social Impact Model
Iterative Hub and Spoke Architecture Hybrid Model Hybrid Model Inference Engine Inference Engine Pave/CG Mediation Ontology (Inference Engines: Probabilistic Inference (Bayesian networks) Logical Inference (Jena Micro OWL)) Pave/Nexus Mediation Ontology (Inference Engines: Probabilistic Inference (Bayesian networks) Logical Inference (Jena Micro OWL)) Updated Indicators Pave Hub Ontology Updated Indicators CG Move CG Adjudication Nexus Adjudication Nexus Move CG Ontology (Cultural Geography Model) Nexus Ontology (Nexus Model) Legend: Input / Output Ontology Social Impact Model
TWG 2010 Probabilistic Ontologies CG ontology. Defines CG moves. Nexus ontology. Defines Nexus moves. PAVE ontology. Hub ontology for model. Contains PAVE moves and role player strategies, goals, and decision points. PAVE CG Mediation ontology. Performs dynamic translation of PAVE tasks to CG moves. PAVE Nexus Mediation ontology. Performs dynamic translation of PAVE tasks to Nexus moves. Tactical Wargame 2010 ontology. Maintains states of the automated role player, such as OAB level/popularity and state of individual move. Multi-Resolutional Bayesian ontology. Defines the macro and micro agents that are used to integrate multi-resolutional models. TEO ontology. Defines events and outcomes. Design of Experiment (DOE) ontology. Abstracts the concept of strategies, goals, and decision points in a doctrinal manner. ProbOnt ontology. Holds the representation of the Bayesian networks that determine selection of events and outcomes. Pakaf ontology. Holds the moves to the Helmand/PAKAF scenario. PakafCgMediation. Automatically translates CG TWG moves to Helmand/PAKAF moves. Social Impact Model
Probabilistic Ontology Relationships DOE TEO PAVE ProbOnt MultiResolutionalBayes PaveCgMediation ProbOnt MultiResolutionalBayes PaveNexusMediation CG Nexus Inheritance Hierarchy and Relationship Structure of TWG Probabilistic Ontologies Social Impact Model
Implementation of Event Probabilities The study team implemented a probabilistic translation from the moves of one model to another using Bayesian networks, from data on event likelihoods. SIMmiddleware added Bayesian Networks, such as the one below, directly to the probabilistic ontology representation. Social Impact Model
The Hub Ontology Contains Study Specific Behaviors Social Impact Model
Crisp Mediation Ontologies Define Exact Translation Between the Hub and a Model Cordon and Search Event Translates into ISAF Attacks Taliban Event if a Leader is Detained Social Impact Model
Unique Implementation of Probability in Ontologies Probabilities are folded into the ontology itself in the form of Macro and Micro Agents Macro Agents describe a group statistically Micro Agents are individuals that may be generated from or may generate statistics Micro Macro Agents enable Multi-resolutional model integration Higher Level Models need only know trends in lower level models as opposed to details Lower Level Models may be run multiple times to get trends, or trends may be taken over multiple individuals Social Impact Model
Macro Agent Defines Chances of Events that Determine an Exact Translation Social Impact Model
Each Distribution Contains Probabilities from the Bayesian Networks Social Impact Model
Micro-Agent Cordon and Search Event without a Detaining of Taliban Leader, is not an attack Social Impact Model
Micro-Agent Cordon and Search with a Detained Leader is interpreted as an Attack on the Taliban Social Impact Model
Preserving the Probabilistic Relations Between Models is important for Validation In risk-based analysis, we want to track the chance that events will happen, proportionately filling in all possibilities to explore the state space Once data is aligned correctly through the probabilistic ontology, it can be compared against other aligned data Markov Processes express the output data over time and over multiple runs Markov Processes can express what actually happened in the real world Probabilistic Distance of Markov Processes can find how close the patterns in the simulation are to what happened in the real world Social Impact Model
Scenario 1 (TWG 2010). Blue Moves/Green Popularity Level of Violence (blue player actions) K = Kinetic M = Medium Kinetic N = Non-Kinetic Popularity of Green P = Popular U = Unpopular Level of Violence Kinetic Medium Kinetic Non-Kinetic .50 .43 .10 K P M P N P Start Scenario .14 .10 .20 .07 .70 Unpopular Popular Level of Green Support .36 .50 .40 .10 .29 .14 K U M U N U .21 .40 .36 11/8/2018 Social Impact Model
Scenario 2 (Excursion). Blue Moves/Green Popularity Level of Violence (blue player actions) K = Kinetic M = Medium Kinetic N = Non-Kinetic Popularity of Green P = Popular U = Unpopular Level of Violence Kinetic Medium Kinetic Non-Kinetic .14 .10 K P M P N P Start Scenario .10 .50 .14 .28 .70 Unpopular Popular Level of Green Support .50 .14 .10 .71 .14 K U M U N U .57 .14 .28 .43 Probabilistic Distance from Scenario 1: 0.09 11/8/2018 Social Impact Model
Validation – Markov Processes from Model and Real World Popularity of Green P = Popular U = Unpopular Level of Violence K = Kinetic M = Medium Kinetic N = Non-Kinetic State Definition Model State Transitions Real World (Afghan Nationwide Quarterly Assessment Review) Apply Kullback-Leibler (KL) divergence to measure probabilistic distance between Markov processes: Score of 0 means exact same Markov process. Score of 1 means the most different Markov process possible. Example above generated KL score = 0.21. Social Impact Model
Summary Social Impact Model
Validation depends on Proper Integration Probabilistic Ontologies can Translate between Models, keeping track of uncertainty. Probabilistic Ontologies can address Multi-Resolutional Model Translation. Once translated, models are aligned. Once aligned, they can express output probabilities and can be aligned with real world data for probabilistic comparison Markov Processes express the data dynamically for comparison at the level of patterns. Social Impact Model
Questions and Comments POC: Deborah Duong dduong@agentBasedLearningSystems.com Social Impact Model