Application of Artificial Intelligence (AI) Programming Techniques to Tactical Guidance for

John W. McManus

and

Kenneth H. Goodrich

NASA Langley Research Center

Mail Stop 4

Hampton, Virginia 23665-5225(804)8-4037/(804)8-4009

AIAA Guidance, Navigation, and Control Conference

August 14-16, 19Boston, Massachusetts

Autonomous Systems and Mission Planning

ABSTRACT

A research program investigating the use of Artificial Intelligence (AI) techniques to aidin the development of a Tactical Decision Generator (TDG) for Within-Visual-Range (WVR)air combat engagements is discussed. The application of AI methods for development andimplementation of the TDG is presented. The history of the Adaptive Maneuvering Logic(AML) program is traced and current versions of the AML program are compared and

contrasted with the TDG system. The Knowledge-Based Systems (KBS) used by the TDG toaid in the decision-making process are outlined in detail and example rules are presented. Theresults of tests to evaluate the performance of the TDG versus a version of AML and versushuman pilots in the Langley Differential Maneuvering Simulator (DMS) are presented. Todate, these results have shown significant performance gains in one-versus-one air combatengagements, and the AI-based TDG software has proven to be much easier to modify than theupdated FORTRAN AML programs.

INTRODUCTION

The development of all-aspect and \"fire and forget\" weapons has increased thecomplexity of the air-to-air combat environment. Modern sensors provide critical tacticalinformation to the aircraft a range and precision that were impossible 20 years ago. Thisincreased complexity, combined with the expanded capabilities of high performance aircraft,has changed the future of air combat engagements. The need for a modern, realistic air combatsimulation that can be used to evaluate the current and future air combat environment has beenwell documented [Burgin 1975, 1986, 1988; Hankins 1979]. Existing tools such as theAdaptive Maneuvering Logic program (AML) [Burgin 1975, 1986, 1988], TAC Brawler[Kerchner 1985], and AASPEM have generally centered their efforts on the development andrefinement of high-fidelity aircraft dynamics modeling techniques and not on the development

and refinement of tactical decision generation logic for WVR engagements. In support of thestudy of superagile aircraft at Langley Research Center (LaRC) a Tactical Guidance Researchand Evaluation System (TGRES, pronounced \"tigress\") is being developed [Goodrich 19].

Figure 1. TGRES SYSTEM.TGRES DESCRIPTION

The TGRES system, shown in figure 1, provides a means by which researchers candevelop and evaluate, in a tactically significant environment, various systems for high

performance aircraft. While TGRES is aimed specifically at the development and evaluation ofmaneuvering strategies and advanced guidance/control systems for superagile aircraft,TGRES's modularity will make it easily adaptable to the analysis of other types of aircraftsystems. TGRES is composed of three main elements--the TDG, the Tactical ManeuverSimulator (TMS) , and the DMS.

The TDG is a knowledge-based guidance system designed to provide insight into thetactical benefits and costs of enhanced aircraft controllability and maneuverability throughout anexpanded flight envelope (i.e. superagility). The two remaining elements of TGRES, the TMSand the DMS, provide simulation environments in which the TDG is exercised. The TMSsimulation environment was developed using conventional computer languages on aVAXStation 3200. The TDG was developed on a Symbolics 3650 workstation. Theseparation of the aircraft simulation and decision logic components allows each module to bedeveloped using hardware and programming techniques specifically designed for its function.This separation of tasks also increases the efficiency of the simulation by allowing someparallel processing. The two processes are executed as co-tasks and communicate via anEtherNet connection. (See fig. 2.)

SYMBOLICS 3650TACTICALDECISIONGENERATORLOGICVAXSTATION 3200CURRENTAIRCRAFTDYNAMICSUSER INTERFACEANDENGAGEMENTREPLAY SYSTEM EtherNetlinkFigure 2. CURRENT HARDWARE CONFIGURATION.

The user interface system consists of a color graphics package designed to replayboth TMS and DMS engagements, and a mouse sensitive representation of the TDG aircraftand its basic systems that allows the user to interact with the TDG aircraft during the

execution of TMS runs. The Engagement Replay System (ERS) software is available for aVAX color workstation and a Symbolics color workstation. The ERS display, shown infigure 3, displays the two aircraft on a three-dimensional axis and has dedicated windowsused to display several aircraft variables including the thrust, Mach number, and deviationangles of the two aircraft. The viewing angle for each engagement can be rotated 360°around both the X and Z axis to provide the most information to the user. The interactiveTMS display includes a graphical representation of the TDG aircraft's major systems suchas engines, offensive and defensive systems, and a system status display. During thesimulation run the user can enable and/or disable the aircraft's systems using the mousesensitive display and evaluate how the changes effect the TDG's decision generationprocess.

Figure 3. ERS DISPLAY.

The final element of TGRES is the Differential Maneuvering Simulator. The DMS consistsof two 40' diameter domes located at Langley. The facility is intended for the real-timesimulation of engagements between piloted aircraft. By using the TDG to drive one of theairplanes, it is possible to test the TDG against a human opponent. This feature allows theguidance logic to be evaluated against an unpredictable and adaptive opponent. A thirddome (20' in diameter) is being added to the DMS facility. This addition will allow theguidance logic to be evaluated in one-versus-two or two-versus-one scenarios, furtherenhancing the tactical capability of the DMS environment.

THE AML PROGRAM

The TDG is being developed as a KBS incorporating some of the features first outlinedin the AML program [Burgin 1975, Hankins 1979]. The AML program was selected as abaseline for several reasons, including its past performance as a real-time WVR tacticaladversary in the Langley DMS and the modular design of the FORTRAN source code. Thetactical decision generation method developed for the original AML program, outlined in figure4, is a unique approach that attempts to model the goal-seeking behavior of a pilot by mappingthe physical situation between the two aircraft into a finite, abstract situation space. A set of thethree basic control variables (bank angle, load factor, and thrust) can be determined to

maximize some performance index in the situation space [Burgin 76]. Each triplet of control

variables defines an \"elemental maneuver,\" and a sequence of these elemental maneuvers mayform classical or \"text book\" air combat maneuvers.

Predict Opponent'sState at t + ∆secRelative Geom.(X, Y, Z)BestManeuverControlCommandsEqs. ofMotionElemental ManeuversManeuver #1Maneuver #2•••Maneuver #nScore ManeuversFigure 4. HOW AML WORKS.

Although the logic and geometry used by AML to make tactical decisions is complex,the basic concepts it uses are simple. At each decision interval, the \"attacking\" aircraft predictsthe future position and velocity of its opponent using a curve-fitting algorithm and past knownpositions of the opponent. The attacker then uses a set of elemental maneuvers (describedabove) to predict a set of positions that it can reach from its current state.

The AML program forms a \"situation state vector\" for each trial maneuver evaluated.The vector is used to represent the responses to a set of questions about the current situation.Figure 5 shows the binary scoring method (0 = NO, 1 = YES) used to determine the value ofeach each cell in the vector.

This vector is multiplied by a \"scoring weight\" vector to form a scalar product that representsthe situation space value for the current maneuver. A detailed description of the trial maneuvergeneration and scoring process and an explanation of how the scoring weights have evolvedcan be found in [Burgin 1988]. The questions used to form the situation state vector wereobtained from several sources including air combat maneuvering manuals, interviews withfighter pilots, and detailed analysis of the original DMS engagements. In the original versionof AML, each question had a positive, non-zero weight. The questions were formulated sothat a \"YES\" answer reflects a favorable condition, increasing the score for the maneuver. It isimportant to note that in the original AML research \"no systematic investigation was made tooptimize these weight factors; they were usually all set to one.\" The early AML versions[Burgin 1975; Hankins 1979] were designed to perform as a conservative opponent. Thescoring rules rated offense and defense evenly and risked giving up some positional advantageto the opponent only when there was reasonable assurance the attacker would gain at least asmuch offsetting advantage. This conservative approach may be the product of a philosophystated in [Burgin 1988],

\"The objective of the decision-making process is to derive maneuvers which will bringone's own weapons to bear on the target while at the same time minimizing exposure tothe other side's weapons.\"

This is a one-dimensional approach to the problem. It outlines a logic that handles only theneutral and aggressive cases effectively and does not recognize that there are several Modes ofOperation (MO), outlined in figure 6, that a pilot may use during an engagement. In manysituations when the opponent has a distinct positional advantage, the AML aircraft will perform\"kamikaze\" maneuvers, giving up one or more clear shots to the opponent while it maneuversto a position of \"advantage.\" In these situations the AML aircraft would not survive to exploitthe positional advantage, having been \"killed\" while obtaining it.

••••AGGRESSIVEDEFENSIVENEUTRALEVASIVE•••EVADING OPPONENTS' \"LOCK AND FIRE\"EVADING MISSILE (AAM & SAM)GROUND / STALL EVASIONFigure 6. TDG MODES OF OPERATION.

The existing trial maneuver versions of AML do a good job of getting behind an

opponent, but due to the grain of the trial maneuvers, lack the ability to fine-track the opponent.Several changes were made to the AML program [Burgin 1988] to address this problem. Therequirement that only the opponents positional data be passed to the algorithm was relaxed and\"complete and accurate information about the the opponent's past and present states\" is nowprovided. The 1986 version of the AML program, AML86 [Burgin 1988], also made severalmajor changes to the tactical decision generation process, abandoning the trial maneuverconcept for a rule-based approach and a set of canned \"Basic Fighter Maneuvers.\" [Burgin1988] contains an extensive history of the \"trial maneuver\" concept and a description of howthe new rule-based version of the program, AML68, was developed. A \"pointing\" controlsystem was also developed to aid the fine-tracking process. The pointing control systemdirectly commands roll and pitch rates to point the aircraft's longitudinal axis at the opponent.AML86 is a first step towards a multi-dimensional approach and is similar to the decision logicincorporated in the TDG.

THE DEVELOPMENT OF THE TDG SYSTEM

The development of the TDG has been a multi-stage process using the COSMIC version ofAML as a starting point. The COSMIC version of AML was updated by DynamicsEngineering Incorporated (DEI) while under contract to NASA Langley. This version ofAML, (DEI-AML), has a scoring module that uses a set of 15 binary questions and a fixed setof weights to evaluate the trial maneuvers. DEI installed aerodynamic data and enginecharacteristics provided by the Aircraft Guidance and Controls Branch (AGCB) into the AMLdata tables and made all changes to the AML software outlined in [Burgin, 1986]. DEI-AMLwas tested by AGCB to insure symmetry of the engagements given symmetric initial

conditions. During the testing process several software bugs were found and corrected. A fulldescription of the bugs and corrections are outlined in [McManus 19]. The resulting code,dubbed AML´, was again tested for symmetry and a DMS ready version of the code, DMS-AML´ was prepared. AML´ and a DMS ready version, DMS-AML´, are being used as thebaseline during development of the TDG system.

Predict Opponent'sState at t + ∆secRelative Geom.(X, Y, Z)Perform SituationAssessmentBestManeuverControlCommandsEqs. ofMotionActive Throttler ControllerScore ManeuversElemental ManeuversGuidance AlgorithmsFigure 7. HOW TDG WORKS.

The TDG system, outlined in figure 7, currently uses the trial maneuver conceptoutlined in the AML program with several extensions. The original set of five to nine trialmaneuvers has been expanded to include over 40 trial maneuvers. Although this is a \"bruteforce\" solution the new trial maneuvers allow the TDG to perform target tracking more

effectively and improve the system's overall performance. The TDG uses an object-orientedprogramming approach to represent each aircraft and the current state of its offensive systems,defensive systems, and engines. This information is used to help guide the TDG's reasoningprocess. The original FORTRAN AML throttle controller and the maneuver scoring moduleshave been redesigned using a rule-based programming approach and ported to the AIworkstation. Examples of rules for each of the KBS modules are shown in figure 8.

((AND (EQUAL (GET-MISSION PALADIN) *AGGRESSIVE*)(EQUAL (GET-POSITION PALADIN) *NEUTRAL*)) ((SETF (GET-MODE PALADIN) *AGGRESSIVE*) (AGGRESSIVE-WEIGHT)))EXAMPLE MODE SELECTION RULE.((AND I-SEE-HIM I-CAN-FIRE HE-CANT-FIRE (<= RANGE 12500)) ((SETF THROTTLE 0.94)) )EXAMPLE THROTTLE CONTROL RULE.((OR (≤ (ABS HIM-UNABLE-TO-FIRE) (I-CAN-FIRE))(AND (> HIM-UNABLE-TO-FIRE O.O) (≥ I-CAN-FIRE 0.0) )(= GUNA 1.0)(= ALLA 1.0)(= HEATA 1.0)) (SETF (GET-POSITION PALADIN) *AGGRESSIVE*))EXAMPLE SITUATION ASSESSMENT RULE.Figure 8. Example Rules.KBS MODULES OF THE TDG

The TDG system has a knowledge-based Situation Assessment (SA) module that isexecuted at each decision interval before the trial maneuvers are evaluated. The SA module isused to determine the TDG's current MO. The SA is executed at each interval, before the

maneuver scoring module, and determines the TDG's MO. This determination is based on theTDG's current mission, the current state of the aircraft's systems, the relative geometry

between the aircraft and its opponent, and the opponent's instantaneous-intent (*in-int*). Eachof the modes shown in figure 6 has a unique set of scoring weights and a decision intervalassociated with it. The weights for each mode have been adjusted during the design and testingprocess to maximize the TDG's performance in that mode. Test results have shown that ashort decision interval, (0.5 sec.), improves the TDG's fine-tracking performance. The sameshort decision interval results in a \"thrashing\" motion in neutral situations resulting in degradedsystem performance. A longer decision interval, (1.0 sec.), is used in neutral situations. Theopponent's *in-int* is defined to be an estimation of the opponent's intent at the current pointin time based on available sensor, positional, and geometric data. Currently, there is no attemptto use a history of *in-int* to derive a long-term opponent intent. The flexibility provided bythe use of MO's allows the system to more closely model the pilots changing strategies duringthe engagement. The COSMIC version of AML, and most AML variations before AML86, donot have the ability to change their decision generation strategy based on the changingenvironment. The TDG Scoring Module (SM) is a KBS that uses a set of 17 fuzzy logic

questions with responses ranging from [0 = NO, ..., 1.0 = YES], (fig. 8), and the set of

mode-specific scoring weights selected by the SA module to score each of the trial maneuvers.A rule-based active Throttle Controller (TC) has been developed to replace the existing throttlecontrol subroutine. The TC is called at the start of each decision interval and can set the throttleat any position from idle to full afterburner [0, .., 2]. The logic for the existing AML throttlecontrol subroutine had only three positions (0 = idle, 1 = military, 2 = full afterburner) and hadbeen turned off in the COSMIC version--all engagements were being flown with the throttle setat full afterburner.

027090TDG180027090AML´180Figure 10. SET OF INITIAL CONDITIONS.

A statistics module is used to calculate the amount of time that each aircraft has its weaponslocked on its opponent and the deviation angle and angle-off. The Line-Of-Sight (LOS) vectoris defined as the vector between ownship c.g. and opponent's c.g. The Line-Of-Sight (LOS)angle is defined as the angle between the LOS vector and ownship body x-axis; the deviationangle is defined as the angle between the LOS vector and ownship velocity vector; and theangle-off is defined as the angle between the LOS vector and opponent's velocity vector (fig.10).

LINE OF SIGHT VECTORANGLE OFFOPPONENT VELOCITY VECTORAML AIRPLANEDEVIATION ANGLEOWNSHIPVELOCITY VECTORTDG AIRPLANEFigure 11. ANGLE DEFINITIONS.

The weapons cones used represent a generic all-aspect missile, a generic tail-aspect missile,and a 20 mm cannon (fig. 11). Four metrics are currently used to evaluate each engagement.The first metric is calculated every second and computes the total time that each airplane has itsweapons locked on the opponent, the probability that the shot will hit, the distance between theopponents, the angle-off, and the deviation angle. The results are printed in a table format atthe completion of each run. The second metric computes a Probability of Survival (PS) usingthe data computed by the first metric. The missiles are treated as a limited resource and aprobability to hit of 0.65 is required to launch the first missile. The firing threshold increasesby 0.05 for each missile launched, and all missiles are required to complete their flight to thetarget before the next missile is fired. The third scoring metric attempts to determine a LethalTime (LT) value for each engagement. The LT value for a run is equal to ((TDG gun time -AML gun time) / 2) + 2 * (TDG tail-aspect time - AML tail-aspect time) + (TDG all-aspect time- AML all-aspect time). A positive LT value shows TDG with an advantage, a negative LTshows AML with an advantage. The fourth metric is Time on Offense (TOF). TOF is the sumof all weapons lock time for each each airplane. ∆TOF is computed as TDG TOF minus AMLTOF.

5° GUN CONE. RANGE 0 TO 5000 FT. TDG- AML (time on offense)2520∆ TOF(seconds)151050-5-10Run Number.Figure 13. ∆TOF FOR SET OF ENGAGEMENTS.

TEST RESULTS

A set of nine engagements presented in [Eidetics ] were used to compare the

performance of the TDG system with the performance of the AML´ in the lab, and againstpilots in the DMS. AML´ was used as the A airplane in both sets of lab test engagements, andthe human pilot flew the A airplane during the DMS runs. Airplanes with identical

performance characteristics were used in both the DMS and the lab. The set of nine initialinitial conditions, fig.13, favor the A airplane.

1AIRCRAFT B.423562 NM SEPARATION.0 KTS AIRSPEED.15,000 ALTITUDE78AIRCRAFT A.9Figure 14. EIDETICS INITIAL CONDITIONS.

The B airplane has five neutral starting positions, runs 3, 5, 6, 7, and 9; one offensive startingposition, run 8; and 3 defensive starting positions, runs 1, 2, and 4. There is a 2-nautical mileseparation between the opponents and each airplane is at an initial altitude of 15,000 feet and aninitial airspeed of 0 knots. All of the engagements were run for 60 seconds. The scoringmetric used was an Overall Exchange Ratio (OER), defined as the # of A killed / the # of Bkilled. The Eidetics study was conducted using a modified version of the AASPEM programand produced an OER of ≈ 0.72. The OER was less than 1.0 due to the use of a non-symmetric set of initial conditions. In the first set of engagements the AML´ program wasflown against itself and the produced an OER of 0.75.

141210Time inSeconds.8201234567Run Number.TOF TDGTOF AMLFigure 15. AML´ vs AML´ TOF.

In the second set of engagements the TDG was used to control the B airplane and achieved anOER of 1.50, a 100 percent improvement. The test results, (figs. 14, 15), clearly show thesuperior performance of the TDG system. It is also interesting to note that the maximum OEREidetics achieved by modifying aircraft performance characteristics was ≈ 0.85 [Eidetics19].

16141210Time inSeconds.8201234567Run Number.TOF tdgTOF amlFigure 16. TDG vs AML´ TOF.

The DMS runs were conducted using the research pilot with the most DMS flight time againstthe TDG-DMS as the opponent. The pilot flew against the set of initial conditions three times,providing a total of 27 runs. TOF data for the DMS runs is not available at this time. TheOER for the set of 27 runs was 0.83. As stated earlier, studies done in the lab have shown thatthe reduced set of trial maneuvers used by DMS-TDG cannot fine track an opponent as

effectively as the expanded set used by the TDG. The reduced set of trial maneuvers used byDMS-TDG may account for most of the performance difference between the TDG and DMS-TDG.

FUTURE WORK

Several enhancements to the existing TDG system are planned. The maneuver selectionlogic will be expanded to replace the use of the trial maneuvers for modes of operation whereconventional guidance algorithms provide better performance. This change to the logic andselection module will improve the TDG's ability to track its opponent. Initial lab results haveshown that the development of mode-specific maneuver sets will increase system efficiency byreducing the number of maneuvers evaluated for some MO's. The development of logic fortwo-vs-one engagements is underway. The third aircraft will be dynamically allocated to eitherthe TDG or the opponent at the start of each run. This feature will allow researchers to evaluatethe TDG in both two-vs-one and one-vs-two engagements. A system for connecting theSymbolics workstation directly to the DMS real-time computing facilities is also being

investigated. The development of such a link would allow the full TDG system to be tested inthe DMS against human pilots.

The TGRES system presents an excellent opportunity to evaluate the use of AIprogramming techniques and knowledge-based systems in a real-time environment. It alsoclearly shows that the maneuver selection and scoring techniques developed in the late 1960'sand early 1970's cannot perform well in the modern tactical environment and are not wellsuited for evaluating agile aircraft. Figure 16 shows many of the changes in the tactical andsimulation environments since the original AML tactical decision generation logic wasdeveloped. The use of KBS and AI programming techniques in developing the TDG hasallowed a complex tactical decision generation system to be developed that addresses themodern combat environment and agile aircraft in a clear and concise manner.

1968HEAT SEEKING WEAPONSDOMINATE TACTICALSITUATIONLIMITED COMPUTING ANDMODELING RESOURCES.SHORT-RANGE RADAR.SHORT-RANGE WEAPONS.1 vs 119ALL-ASPECT WEAPONS DOMINATETACTICAL SITUATION. (LONGERRANGE, FIRE AND FORGET,....)BETTER COMPUTING ANDMODELING RESOURCES.LONG-RANGE RADARLONG-RANGE WEAPONS.2 vs 1, M vs NSUPERMANEUVERABLE AIRCRAFT,POINT AND SHOOT CAPABILITYFigure 17. 1968 AML vs 19 TDG.

CONCLUDING REMARKS

A KBS TDG is being developed to study WVR air combat engagements. The systemincorporates modern airplane simulation techniques, sensors, and weapons systems. Thesystem was developed using several concepts first outlined in the AML program originallydeveloped for use in the LaRC DMS. An updated AML system is being used as a baseline toassess the functional and performance tradeoffs between a conventionally coded system and theAI-based system. Test results have shown that the AI-based TDG system has performed betterthan AML´ in both the TMS and the DMS. The use of a KBS SA module and MO's allows theTDG to more accurately represent the complex decision making process carried out by a pilot.The use of a more extensive set of trial maneuvers and a KBS TC module allows the TDG tofine track the opponent more effectively than AML´. The KBS decision generation logic hasproved to be much easier to modify than the AML´ FORTRAN source code. The ability tointegrate the TDG into the DMS offers a unique opportunity to evaluate the performance of theAI-based TDG software in a real-time tactical environment against human pilots.

REFERENCES

1.

Burgin, G. H. ; et al. : An Adaptive Maneuvering Logic Computer Program for the

Simulation of One-on-One Air-to-Air Combat. Vol I and II. NASA CR-2582, CR-2583,1975.

Burgin, G. H. : Improvements to the Adaptive Maneuvering Logic Program. NASA CR-3985, 1986.

Burgin, G. H. ; and Sidor, L. B. : Rule-Based Air Combat Simulation. NASA CR-4160,1988.

Hankins III, W. W. : Computer-Automated Opponent for Manned Air-to-Air CombatSimulations. NASA TP-1518, 1979.

2. 3. 4.

5. 6. 7.

Kerchner, R. M. ; et al. : The TAC Brawler Air Combat Simulation Analyst Manual(Revision 3. 0). DSA Report #668.Buttrill, C. S. ; et al. : Draft NASA TM 19.

Taylor, Robert T. ; et al. : Simulated Combat for Advanced Technology Assessments

Utilizing The Adaptive Maneuvering Logic Concepts. NASA Order no. L-24468C, CoastalDynamics Technical Report No. 87-001.

McManus, John W. ; Goodrich, Kenneth H. : Draft NASA TM 19.Goodrich, Kenneth H; McManus John W. :AIAA Paper #...

8.9.