Combine Intent Recognition with Behavior Modeling in Teaching

Competition Military Simulation Platform

Yi Zhang, Shuilin Li, Chuan Ai, Yong Peng

and Kai Xu

College of Systems Engineering, National University of Defense Technology, Changsha 410073, China

Keywords: Simulation and Application, Evolution Behavior Tree, Symbolic Plan Recognition, Decision-Reasoning.

Abstract: Intent recognition refers to obtaining the observations of an agent and then using the observations to reason

its current state and to predict its future actions. Behavior modeling, describing the behavior or performance

of an agent, is an important research area in intent recognition. However, few studies have combined behavior

modeling with intent recognition to investigate its real-world applications. In this paper, we study behavior

modeling for intent recognition for cognitive intelligence, aiming to enhance the situational awareness

capability of AI and expand its applications in multiple fields. Taking the combat environment and tanks as

the research object, based on the behavior tree and SBR recognition algorithm, this paper designs the

framework and experiments for behavior modeling and intent recognition. Firstly, uses the evolution behavior

tree algorithm to autonomously generate the behavior model adapted to the environment. Secondly uses the

SBR algorithm to effectively recognize actions and plan paths of enemy tank to guide self-tank actions in the

TankSimV1.20 simulation platform. The results show that the tank survival rate increases by 80% under the

guidance of the intent recognition results, and the method in this paper can provide effective guidance for the

intent recognition behavior modeling, which has a broad application prospect.

1 INTRODUCTION

Intent recognition, the ability to recognize the

activities, plans and goals of other agents, enables the

observer to reason about the current state of the

recognized agent and predict its future action(Mirsky

et al., n.d.). In practical research, intent recognition

contains three types: goal recognition, plan

recognition, and activity recognition. Among these,

activity recognition is the least abstract level of

inference, goal recognition is the most abstract, and

plan recognition lies somewhere between them.

Behavior modeling, describing the behavior or

performance of an identified person, is an important

research area in intent recognition. Behavior

modeling is mainly divided into cognitive behavior

modeling and physical behavior modeling. Physical

behavior modeling refers to the direct physical

modeling of the external environment. Cognitive

behavior modeling refers to simulating the various

thinking processes of the recognized agent.

In military simulation, behavior trees are often

utilized to guide the Computer-Generated Forces

(CGFs) for simulating combat processes(Fu

Yanchang, 2019). While Behavior Trees can

effectively manage and organize a series of

predefined behavior patterns, the CGFs lack the

ability to flexibly respond and make independent

decisions based on actual situations(Jie Yang, 2021).

If the integration of intent recognition and behavior

trees can be achieved and applied in military

simulation, it would enable CGFs to possess more

Observation

Orientation

Decision

Action

SBR Algorithm EBT Models

Tank AgentTankSimV1.20

Red Tank SBR Algorithm Blue Tank EBT Models

Observation

Plan Library

Current State Qurey(CSQ)

Propagate Up

Reason Plan & Goal

ConMove

Move

ConAttack Attack

Figure 1: The framework of combine intent recognition

with behavior modeling in teaching competition military

simulation platform (TCMSP).

456

Zhang, Y., Li, S., Ai, C., Peng, Y. and Xu, K.

Combine Intent Recognition with Behavior Modeling in Teaching Competition Military Simulation Platform.

DOI: 10.5220/0012869400003758

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH 2024), pages 456-463

ISBN: 978-989-758-708-5; ISSN: 2184-2841

advanced decision-making capabilities, thereby

improving the realism and practicality of military

simulations. This work seeks to combine intent

recognition with behavior modeling in military

simulation platform, and thus allow a decision-

making agent.

Figure 1 shows the framework of this paper. The

circle consisting of four modules formulates a

cognition loop(Xu et al., 2019). Specifically, the

TankSim V1.20 firstly feed the observation into EBT

Models, which is a behavior model that we generate

using the Evolutionary Algorithm. The EBT models

guides blue tank actions, making them more closely

resemble real-world military forces(Jie Yang, 2021).

Then the blue tank makes real-time decisions. Finally,

the red tank makes decisions based on actions of blue

tank using the SBR algorithm(Avrahami-Zilberbrand

& Kaminka, 2005).

The rest of this paper is arranged as follows:

Section 2 introduces related work. Section 3

introduces the algorithms we used. Section 4

introduces the experiments and results of this work.

Section 5 introduces conclusions and future work.

2 RELATED WORK

2.1 Intent Recognition Algorithms

Since the formal definition of planning recognition

process was proposed in 1978(Schmidt et al., 1978),

scholars have started to study methods for plan

recognition. A method of generalized planning

recognition was proposed in 1986(Kautz & Allen,

1986), which describes task plan recognition with a

planning graph, represents the decomposition of the

task with the vertices of the graph, and proposes to

use the graph overlay for the solution of the problem,

which to some extent laid the foundation for

subsequent research. This method, although efficient,

assumes that the top-level goal of plan is unique and

does not consider the different priori probabilities of

different goals.

The Symbolic Plan Recognition (SBR) method

was proposed in 2005 (Avrahami-Zilberbrand &

Kaminka, 2005), which efficiently implements plan

recognition using tagging and back propagating, and

can quickly give partial solutions thus applying to

multiple aspects, but the efficiency decreases when

multiple plans are run concurrently. Further, the

authors proposed Utility-based Plan Recognition

(UPR)(Avrahami-Zilberbrand & Kaminka, 2007),

which can recognize multiple plans in overlapping

and interleaved contexts. The SBR family of

algorithms runs efficiently and can produce results in

each time and is usually used as a frequent choice of

recognition method by researchers.

A probabilistic planning recognition algorithm

based on planning tree grammar was proposed in

2009(Geib, 2009), which regarded plan recognition

as the parsing of the grammar tree, which effectively

solved the plan recognition in the case of multiple

concurrent plans, but it needs to construct a complete

parsing set. In the same period, a Planning

Recognition as Planning (PRaP) approach was

proposed in 2009(Ramírez & Geffner, 2009), which

used planning techniques to solve the goal

recognition problem by comparing the marginal cost,

which is the difference between the consistency of a

given observation and the best plan, between different

plans for the same goal through multiple invocations

of the AI planning system. This method is limited by

the fact that it can only reason about one plan at a time

and is computationally expensive.

Since then, attention has been paid to improving

the performance of the recognizer. A cost-based goal

recognition method that improves the speed of the

recognizer compared to PRaP, but is limited to the

path planning domain(Masters & Sardina, 2017). A

method of sampling the parse space using Monte-

Carlo tree search can significantly improves the speed

of solution compared to full parse (Kantharaju et al.,

2019).

2.2 Behavior Modelling Algorithms

Currently, commonly used behavior modeling

methods include, but are not limited to, the behavior

tree (BT), the finite state machines (FSM), and the

dynamic script (DS).

A rule-based behavior decision-making algorithm

for unidirectional two-channels by combining fuzzy

inference with a finite state machine was proposed in

2023(WANG Liang et al., 2023). A vehicle-level

expected functional safety hazard recognition method

based on a model of finite state machine was

proposed in 2023(XIONG Lu et al., 2023). Due to the

special structural characteristics, finite state machines

can only save a finite number of steps of state transfer,

so it is difficult for the FSM system to monitor the

historical execution flow of the transfer from the

initial state to the final state, and vice versa.

Dynamic Scripts (DS) is a reinforcement learning

technique based on rule scripts proposed in

2006(Spronck et al., 2006). In 2015, an Evolution

Dynamic Script (EDS), which embeds an evolution

approach to discovering new rules during DS learning

Combine Intent Recognition with Behavior Modeling in Teaching Competition Military Simulation Platform

457

was proposed in 2015(Kop et al., 2015). The

representation of rule scripts improves the

comprehensibility of the model, but the quality of the

rule base greatly affects the quality of the generated

model.

As a mathematical model, behavior tree describes

the transfer between finite tasks in a modular way,

which allows the creation of complex tasks with

simple tasks without considering the execution

process of the underlying simple tasks(XIAO Zichao,

2020), and is commonly used for the execution of

tasks in fields such as computer science, control

systems, robotics, and video games.

A behavior tree-based CGF behavior modelling

was conducted in 2019(Fu Yanchang, 2019), which

decomposed complex mission objectives into a

hierarchical structure represented by behavior

subtrees, and optimized the behavior tree by

introducing machine learning methods to assist the

modeling. In 2018, an integrated framework

including a case-based reasoning evolution behavior

tree, and reinforcement learning to facilitate CGF

autonomous behavior modelling was

proposed(Zhang et al., 2018).

Based on the behavior tree, a paper designs an

autonomous generation framework of CGF decision

model, which used an evolution behavior tree

algorithm based on static constraints to efficiently

generate behavior tree decision model according to

expert domain knowledge(Jie Yang, 2021). Since this

paper is very inspiring, it will be used as an important

paper for our reference.

Most of these scholars are limited to one type of

research direction. In fact, it is the merging of

behavior modeling and intent recognition research

that can advance the development of cognitive

intelligence in AI. In this paper, behavior modeling

and intent recognition are merged and simulated and

experimented in the TSMCP, which can provide more

comprehensive and accurate solutions for practical

applications and promote the development of related

fields.

3 BACKGROUND APPROACHES

This section provides some background on simulation

environment, behavior tree and SBR. It includes

behavior tree basics and SBR-related representations.

3.1 Simulation Environment

TankSimV1.20 simulation environment is a TCMSP

developed by the College of Systems Engineering,

National University of Defense Technology in 2020.

In this environment, the user can generate tank

platoons and assign behavior tree rules to each tank.

Under the guidance of the behavior tree rules, the

tanks can roam around the map and complete the

annihilation mission. Figure 2 shows the simulation

environment.

Figure 2: Simulation Environment TankSim V1.20.

The environment is set as follows: based on the

TankSimV1.20 simulation environment, the tank

platoons of both sides, red and blue, meet each other

on a certain plain terrain, which consists of 14*14

squares. Both sides need to carry out annihilation

tasks within a certain period. In this scenario, both

sides carry a certain number of missiles to attack each

other and can open the shield to defend against

missiles. Both tanks are equipped with a series of

sensors to detect each element on the map. The tank

that defeats the opponent or has the most resources

remaining in a limited time wins.

3.2 Evolution Behavior Tree

Behavior tree is used to control the actions of tanks,

writing different behavior tree rules for tanks can

make tanks realize different actions. Tanks with

different behavior trees achieve different results in the

competition, for example, tanks that only attack will

not avoid the missile attack of enemy tank and will

suffer heavy losses in the ambush; tanks that only

escape will not attack the enemy tanks and will not

win the war. Therefore, the tank behavior tree rule

determines the victory or defeat of tanks in combat

simulation.

Evolution algorithms, inspired by the evolution

mechanisms of biological populations in nature, can

search for optimal solutions to optimization problems

and are known as intelligent methods, highly robust

and adaptive, as well as being implicitly parallel and

self-learning. Inspired by this paper(Jie Yang, 2021),

SIMULTECH 2024 - 14th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

458

we used the proposed evolution behavior tree

algorithm to generate an agent behavior model,

shown in Algorithm 1.

Data: Parameters

Output: Best behavior tree model

for i ∈ epoch

Selection

Generate New

Crossove

Mutation

end

return Behavior Tree Model

Algorithm 1: Evolution Behavior Algorithm.

3.3 SBR Algorithm

Symbolic Plan Recognition (SBR) was proposed in

2005 inspired by automated planning methods, which

are more efficient at runtime at the expense of

solution completeness and can produce partial results

in each time. Therefore, this method is chosen for this

paper for intelligent body intent recognition

(Avrahami-Zilberbrand & Kaminka, 2005).

The SBR algorithm, shown in Algorithm 2,

consists of the following three steps: describing the

behavior of the intelligent body using the Plan

Library (PL); efficiently matching the observations

with the plans in the PL(CurrentStateQuery); and

inferring the plan path (PropagateUp).

Data: Observations, PL

Output: Plan Path

for o ∈ Observations

CurrentStateQuer

()

Propa

ateUp()

end

return Plan Path

Algorithm 2: Symbolic Plan Recognition.

4 EXPERIMENTAL STUDY

In the experimental study, train against manually set

behavior tree scripts to generate superior behavioral

tree rules by EBT algorithm. The generated behavior

tree was used to guide the blue decision and the

simple behavioral tree was used to guide the red

decision. The two are played against each other under

the above rules, and the survival rate and the energy

remaining are used as the recognition evaluation

metrics.

4.1 Behavior Tree

Firstly, based on the specific requirements of

TankSimV1.20, we design the specific conditions and

action nodes. Then the script was manually written as

an adversarial script model. Finally, the results were

analysed by evolution behavior tree algorithm.

Set the Base Node of the Behavior Tree

We set the basic nodes of the behavior tree like Figure

3. The meaning of each node in the figure is its

English meaning, where the ellipse represents the

condition, and the square represents the action.

RetreatMove

Attack

Turn

RetreatTurn

RetreatTurm

TurnToFire

MoveToGoal

Figure 3: The behavior tree base nodes.

Set Parameters of EBT Algorithm

The population size is set to 100; the rest of the

parameters are set as listed in the Table 1.

Table 1: The parameters for the evolution behavior tree.

Paramete

Value

Population Size 100

Evolution Epoch 100

Initial minimum len

th 14

Initial maximum len

th 18

Crossover probabilit

0.8

Mutation probabilit

0.05

Selection probabilit

0.1

ew unit probabilit

0.05

Set Confront Script Model

The confront model we used to train the tank shows

in Figure 4.

EBT Generates Model

With a crossover probability of 0.6, the EBT

performance is optimal, and the population has the

highest average fitness value. Under the crossover

probability of 0.5 and 0.6, the EBT population

converges more slowly, and it starts to converge only

40 and 50 generations, respectively. The fitness

Combine Intent Recognition with Behavior Modeling in Teaching Competition Military Simulation Platform

459

MoveToGoal

Attack

RetreatTurn

Attack

TurnToFire

Figure 4: The confront behavior tree nodes.

to more than 0, which indicates that the behavior tree

generated by the EBT under this probability is

sufficiently strong to cope with the environmental

changes. Figure 5 illustrated the fitness curve at

different probabilities.

Figure 5: Fitness curve at different probabilities.

The evolution behavior tree corresponding to the

highest average fitness value is selected as the final

more excellent behavior tree, and the evolution with

the highest average fitness value is the 50th

generation of evolution, with a fitness value of 32.

Figure 6 shows the best behavior tree model of the

tank.

Attack

TurnToFire

MoveToGoal

Turn

Turm

Figure 6: The best behavior tree model of the tank.

4.2 SBR Recognition Effect

This section analyses the recognition effects of two

types of behavior trees given to red tank, one is a

simple behavior tree with only roaming on the map,

and the other is a complex behavior tree with both

roaming and attacking functions. The opposing tank

has complex behavior tree rules generated by EBT.

4.2.1 Simple BT

The roaming simple behavior tree plan library is

shown in Figure 7(Mirsky et al., 2022). From this

plan library, the tank can perform include simple

actions such as steering to turn on the radar, turning

backward, etc. Solid lines represent how each node is

decomposed into other nodes and ﬁnally to actions.

Dashed lines show the order by which the nodes

should appear.

Wander

Turn&OpenRadar

Turn

Radar on

Shields on

Power Set

Forward&ControlRadar

Control Radar

Forward Shields Off

CheckRadarWave

Move Radar on Shields on

Power Set

Figure 7: The simple BT example.

Firstly, to facilitate the analysis, only add the red

and blue sides a tank respectively. Both sides of the

tanks randomly appeared in the map when the

simulation beginning. Because the red side has no

attack ability, the end of the run shows that the blue

tank wins. The recognition effect is as shown Table 2,

which shows SBR has a perfect 100% recognition

rate for simple behavior trees.

Table 2: The simple BT recognition results.

Observation Recognition path Rate

Turn

Turn & Open

Rada

-Turn

100

Forward

Forward & Control Radar-

Forwar

100

Shields off

Forward & Control Radar-

Shields off

100

4.2.2 Complete BT

To verify the correctness of the SBR algorithm in

recognizing complex behavior tree, 10 groups of

experiments are carried out. The initial tank location

was randomized for each set of experiments and 10

steps of the actions of tank were analyzed for each

experiment.

-230

-180

-130

-80

-30

0 20406080100

Fitness

Epoch

Confront Results

CP=0.8 CP=0.7 CP=0.6 CP=0.5

SIMULTECH 2024 - 14th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

460

Right

Retreat

Slide&ShiledsOn

Slide

Turn Left

Shields On

Turn Right

Moveforward

Forward&ShiledsOn

Shields On Moveforward

Backward&ShiledsOn

Shields On

Movebackward

Chase

Turn&OpenRadar

Turn

Radar On

Power Set

Forward

Left

Attack

Fire&ShiledsOn Fire

Shileds On

Slide

Turn

Right

Forward

Left

Turn&Fire

Turn

Right

Fire

Left

Right

Figure 8: The complete BT example.

Since each step of the tank executes three or two

actions of the top-level plan, and there is overlap in

actions between the top-level plans, the tank plan

library can be set to be fully distinguishable or

partially distinguishable. The fully distinguishable

case is like a simple behavior tree, with a recognition

rate of 100%. Therefore, our experiments focus on

analyzing partially distinguishable cases.

Plan library partially distinguishable means the

same action isn’t described with additional distinction

of different top-level plans, e.g., turn right belongs to

the same Wander and Chase top-level plans, but no

additional information is made for each. In this case,

the results of conducting 10 sets of experiments are

shown in Table 3.The experimental results show that

in the partially distinguishable case, the algorithm

recognizes all possible outcomes, and cannot infer the

exact action without additional information.

When the recognized action is Fire, since only the

Attack top-level plan contains it, the recognition rate

is 100%. However, when the recognized action is

Shields on, the recognition algorithm is unable to

determine exactly which plan it is since all three top-

level plans, Attack, Wander and Retreat, contain this

node. Nonetheless, if in practical applications, even if

there are multiple possible recognition results, these

results do not affect the correctness of the final

decision, we can still consider that the algorithm

performs well at the decision level.

4.3 Combination SBR with BT

In this section, the blue tank has the behavior rules

generated by the behavior tree, and the red side has

the simple rules. Use SBR to recognize actions of

blue and guide decision of red making based on the

recognition results. Analyze resource surplus of red

and its survival time.

Table 3: Partially distinguishable experiments.

Group Trial Recognition Actually

1 Attack/Wander/Retrea

Attac

Attack-Fire Attack

Wander/Attack/Chase

Wander

Wander Wander

…

1 Wander/Chase/Retrea

Wande

2Wande

Wande

4 Attack/Wander/Retrea

Attac

5 Attac

-Fire Attac

…

1 Attack/Wander/Retrea

Attac

2 Attac

-Fire Attac

3Wande

Wande

4Wande

Wande

5 Wander/Attack/Chase Wande

…

4.3.1 Compare Survival Time

To compare survival rates, we set up ten sets of

experiments, each comparing decision making

with/without SBR guidance. With SBR guidance, the

red tank could reason the action of blue, and change

self-state to protect self from the annihilate of blue,

while without SBR the red tank could only make

decisions based on the environment.

We set up the initial condition as a small range

annihilation battle, and the blue side will annihilate

the enemy in a limited step length and analyze the

decision-making and survival time of the red side, and

the results show in Figure 9.

The experimental results show that with SBR

guidance, the red tank easily escapes from the blue

tank encirclement, while the survival time is greatly

improved. This result suggests that in a combat

Combine Intent Recognition with Behavior Modeling in Teaching Competition Military Simulation Platform

461

environment where the self-party is at a disadvantage,

it can be assisted in escaping by making a recognition

decision through observation of the enemy.

Figure 9: The loss number and escape time results.

4.3.2 Compare Energy Changes

To compare energy changes, we set up 10 sets of

experiments, each with 60-time stamp data

comparing decisions with/without SBR guidance.

Without setting initial conditions for the tanks, we

analyze the resource changes (in this case energy

changes) of the red tanks as they roam through the

environment, and the results show in Figure 10.

Figure 10: The energy change results.

The experimental results show that with SBR

guidance, the red tank consumes less resources. Since

it can recognize enemy movements and make

decisions, it can reduce the consumption of

unnecessary energy to detect the enemy. This result

suggests that in a combat environment where the

mission of the party is to explore the environment and

avoid the enemy, having SBR guidance reduces

energy consumption and allows the self-party to

explore deeper and find resources.

5 CONCLUSIONS

In this paper, we have combined intent recognition

with behavior modeling in TCMSP. Two methods,

including behavior modeling of EBT, intent

recognition of SBR are used. The framework of

application above is simple but inspiring.

Specifically analyze recognition effects of two

kinds of behavior rules of the tanks: simple roaming

and complex attack. Experiments show that the SBR

algorithm can recognize correctly up to 100% in the

case of fully distinguishable plan library, while in the

partially distinguishable libraries, affected by the

representation of plan library and the number of the

same actions, the average recognition correctness is

70% (in this paper).

At the same time, in the process of simple rules

possessed by the red tank, the results of SBR

recognizing the blue are used to assist the decision-

making of the red side. Experiments show that under

annihilation war, the red side survives longer and

escapes easily; under resource war, the red side

consumes less energy. These results show that

effective recognition of complex behaviors of tank

can provide strong support for intelligent decision-

making.

Though this work is quite suggestive but still

simple, problems are more complex and adversarial

are still open in the future. The proposed approach

could be enhanced by the following future work:

(1) Obtain the PL of the observed Agent

automatically by learning algorithms and LLM.

Solving the PL Representation Problem by a priori

probability.

(2) Apply modeling and recognition methods to

more complex systems like multiagent.

REFERENCES

Avrahami-Zilberbrand, D., & Kaminka, G. (2005, July 30).

Fast and Complete Symbolic Plan Recognition.

International Joint Conference on Artificial

Intelligence.

Avrahami-Zilberbrand, D., & Kaminka, G. (2007, July 22).

Incorporating Observer Biases in Keyhole Plan

Recognition (Efficiently!). AAAI Conference on

Artificial Intelligence.

Fu Yanchang. (2019). Research on CGF Behavior Model

Based on Behavior Tree [M.E., National University of

Defense Technology].

Geib, C. (2009). Delaying Commitment in Plan Recognition

Using Combinatory Categorial Grammars.

Proceedings of the 21st International Joint Conference

on Artificial Intelligence.

LOSE ESCAPE TIME

Annihilate War

No SBR SBR

200

400

600

800

1000

1200

1 163146

Energy

Time Stamp

Energy Change

SBR

No SBR

SIMULTECH 2024 - 14th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

462

Jie Yang. (2021). Research on Generation Method of

Behavior Tree Decision Model for Computer

Generated Forces(CGFs) [M.E., National University

of Defense Technology].

Kantharaju, P., Ontanon, S., & Geib, C. W. (2019). Scaling

up CCG-Based Plan Recognition via Monte-Carlo Tree

Search. 2019 IEEE Conference on Games (CoG), 1–8.

Kautz, H. A., & Allen, J. F. (1986). Generalized plan

recognition. Proceedings of the Fifth AAAI National

Conference on Artificial Intelligence, 32–37.

Kop, R., Toubman, A., Hoogendoorn, M., & Roessingh, J.

J. (2015). Evolutionary Dynamic Scripting: Adaptation

of Expert Rule Bases for Serious Games. Current

Approaches in Applied Artificial Intelligence, 53–62.

Masters, P., & Sardina, S. (2017). Cost-Based Goal

Recognition for Path-Planning. Proceedings of the 16th

Conference on Autonomous Agents and MultiAgent

Systems, 750–758.

Mirsky, R., Galun, R., Gal, K., & Kaminka, G. (2022).

Comparing Plan Recognition Algorithms Through

Standard Plan Libraries. Frontiers in Artificial

Intelligence, 4.

Mirsky, R., Sarah, K., & Christopher, G. (n.d.).

Introduction to Symbolic Plan and Goal Recognition.

Ramírez, M., & Geffner, H. (2009, July 11). Plan

Recognition as Planning. International Joint

Conference on Artificial Intelligence.

Schmidt, C. F., Sridharan, N. S., & Goodson, J. L. (1978).

The plan recognition problem: An intersection of

psychology and artificial intelligence. Artificial

Intelligence, 11(1), 45–83.

Spronck, P., Ponsen, M., Sprinkhuizen-Kuyper, I., &

Postma, E. (2006). Adaptive game AI with dynamic

scripting. Machine Learning, 63(3), 217–248.

WANG Liang, SU Dong-xu, YANG Xing-long, MA Wen-

feng, & WANG Zi-jun. (2023). Behavioral Decision

Based on Fuzzy Inference and Finite State Machine.

Value Engineering, 42(18), 133–135.

XIAO Zichao. (2020). Application Research of Behavior

Tree in Self-driving Behavior Planning Strategy [M.E.,

Lanzhou University].

XIONG Lu, JIA Tong, CHEN Junyi, XING Xingyu, & LI

Bo. (2023). Hazard Identification Method for Safety of

the Intended Functionality Based on Finite State

Machine. JOURNAL OF TONGJI

UNIVERSITY(NATURAL SCIENCE), 51(4), 616–

622.

Xu, K., Zeng, Y., Zhang, Q., Yin, Q., Sun, L., & Xiao, K.

(2019). Online probabilistic goal recognition and its

application in dynamic shortest-path local network

interdiction. Engineering Applications of Artificial

Intelligence, 85, 57–71.

Zhang, Q., Yao, J., Yin, Q., & Zha, Y. (2018). Learning

Behavior Trees for Autonomous Agents with Hybrid

Constraints Evolution. Applied Sciences, 8(7), 1077.

Combine Intent Recognition with Behavior Modeling in Teaching Competition Military Simulation Platform

463