dtank Updated: Exploring Moderated Behavior in a Light-weight Synthetic Environment

Transcription

1 dtank Updated: Exploring Moderated Behavior in a Light-weight Synthetic Environment Frank Ritter 1, Sue Kase 1, Damodar Bhandarkar 1, Bil Lewis 2, Mark Cohen 3 1 Pennsylvania State University College of Information Sciences and Technology 2 Tufts University 3 Lock Haven University The development of this software was supported by ONR (contract N ). This work was also supported by the UK MOD's Analysis, Experimentation and Simulation corporate research programme (Project No: RT/COM/3/006).

2 New dtank v.4.0 Adversarial simulation environment for combat vehicles Light-weight and highly-usable Client-server architecture and socket-based interface Tank agent-type comparisons (Jess, Java, Soar) Programmable vehicle commander behavior based on information gathered from environmental scans Variable battalion configurations Other vehicle and terrain types addable Logs vehicle health and battlefield statistics Ideal for rapid prototyping of situation awareness and moderator-influenced behavior 2

3 User interface User configures battle: Battalions Allied vs. Axis Dimensions of the grid map range 500 to 10,000 m Duration of battle Tank commanders implement battle strategies by executing tactical code Type of tanks being commanded At start of battle, tanks are lined up on opposing sides of the map, then the simulation engine is started 3

4 During a simulation run tanks scan the battlefield acting on the information gathered according to their tactical code Tanks move and projectiles fire similar to laws of physics with damage from hits logged in the server display Summary of damaged and destroyed tanks displays at top of the server display along with expiring battle time in seconds 4

5 Logged data At battle termination, damage and weapon statistics by tank are displayed to active terminal window Nonfatal hits on tanks may damage armor, treads, destroy the radio or give the crew headaches Allied: Sherman4 AFV_is: Operational Damage:0.4 Radio:Nonoperational Tread:Nonoperational Shells_Fired:17 ========== Starting Battle 'null' Combattants: Axis: AlliedCommander vs. Allied: AlliedCommander =========== DumbCommander Allied Sherman NoDisplay 1 DumbCommander Successful_Shots:4 Allied Sherman NoDisplay 1 DumbCommander Allied Sherman NoDisplay 1 SmartCommander Axis Tiger NoDisplay 1 SmartCommander Axis Tiger NoDisplay 1 SmartCommander Axis Tiger NoDisplay 1 Allies: Total:3 Damaged:0 Destroyed:3 Axis: Total:3 Damaged:1 Destroyed:0 Axis: Tiger0 AFV_is:Operational Damage:0.0 Radio:Operational Tread:Operational Shells_Fired:17 Successful_Shots:2 Axis: Tiger1 AFV_is:Operational Damage:0.0 Radio:Operational Tread:Nonoperational Shells_Fired:18 Successful_Shots:1 Axis: Tiger2 AFV_is:Operational Damage:0.0 Radio:Operational Tread:Operational Shells_Fired:12 Successful_Shots:8 Allied: Sherman3 AFV_is:Destroyed Tread:Nonoperational Shells_Fired:3 Successful_Shots:2 Allied: Sherman4 AFV_is:Destroyed Tread:Nonoperational Shells_Fired:0 Successful_Shots:0 Allied: Sherman5 AFV_is:Destroyed Tread:Nonoperational Shells_Fired:0 Successful_Shots:0 ========== Battles over ========== 5

6 Jess built-in tank Built-in Tanks using other architectures The Jess rule-based reasoning engine can be connected to dtank - a common working memory is generated - when the environment is scanned all information gathered is transmitted to the common working memory - based on matched rules, actions can be specified for the Jess Commander Soar built-in tank The connection between Soar and dtank is initialed by creating a Soar Kernel and loading the Soar productions that define a commander s behavior - the productions watch the input link for changes in the environment and respond by placing appropriate actions on the output link which dtank monitors for pending actions Example Soar Commander tactical strategy Soar defines behavior as movement through a problem space Problem space hierarchy for Soar Commander 6

7 Very simple SA Experiment: environmental scanning delay Battle Set (30) 1 2 Team Additional Delay (s) Wins Alert 0 14 Dull 0 13 Alert 0 16 Dull 1 8 Ties sets of 30 battles Commanders engage wander-retreat-attack strategy from previous slide 4 Alert Commanders vs. 4 Dull Commanders increasing delay added to Dull Commanders 3 Alert 0 13 Dull to simulate fatigue or stress 4 Alert Dull 3 2 *Win =least number of destroyed tanks Summary of results Alert Commanders win more battles as Dull Commanders become more passive as their scanning delay is increased (as expected) Set 1: Both commanders scan the environment every 1 second Set 2: Additional scanning delay of 1 second for Dull Commanders Set 3: Additional scanning delay of 2 seconds for Dull Commanders Set 4: Additional scanning delay of 3 seconds for Dull Commanders 7

8 Means analysis Upper plot Mean destroyed tanks per Dull Commander - nearly horizontal line from no-delay scanning state up to 2s delay; then upward trend for delay increase from 2 to 3 s Lower plot Mean success shots for Dull Commanders - steep downward trend from nodelay state to 3s delay; while slight increase visible for Alert Commanders * The environment is sensitive to approximate delays caused by poor SA 8

9 Conclusion & Future dtank A light-weight sandbox for developmental studies Configurable environment Works with Soar, Jess, and Java Supports teamwork through communication between tank agents Sensitive to approximate time scales of cognitive architectures for SA studies eg. CoJack and moderators Many opportunities for improvement More environmental information needs to be accessible to tank agents! 9