Skip to content
Branch: master
Go to file

Latest commit


Failed to load latest commit information.
Latest commit message
Commit time


An Artificial Intelligence simulation engine

Tim Potze
June 4, 2015

The source code of this project can also be found on


  • Introduction
  • The Simulation Engine
    • Controls
    • Namespacing
    • Entities
    • Scripting
    • Services
    • Steering
    • Path Planning
    • Fuzzy Logic
  • Temple of Kotmogu
    • Behavior
    • Controls
    • Fuzzy Logic
  • Conclusion
    • Room for Improvements


For the assignment of Artificial Intelligence for Games I've decided to build a full simulation engine in which simulations can be programmed using a scripting language. The scripting language which can be used is the Pawn Language which has been developed by Thiadmer Riemersma and is written in C. The Pawn language has been used by various gaming related projects, such as the SourceMod and San Andreas Multiplayer game modifications. Since my project is written in C# and the Pawn abstract machine is written in C, I've written a wrapper for it.

The solution found within the source(src) directiory contains three projects:

  • AIWorld: The simulation engine described in this document.
  • amx32: The pawn abstract machine executor.
  • AMXWrapper: A wrapper I've written for the abstract machine executor.

This document consists of two parts. The first part explains the how the simulation engine works. The second part explains how the Temple of Kotmogu simulation works.

The Simulation Engine


By default, a couple of keys have a predefined behavior within the simulation. By pressing the F5 key the simulation is reloaded. Using the arrow keys the camera can be moved. By dragging the mouse while pressing the middle or right mouse button the camera can be rotated. Using the scroll wheel the camera can zoom in and out. By pressing the Tab key, the console can be shown or hidden. By pressing the page up/page down keys or by scrolling while the console is open you can go up and down in the console messages history.


The AIWorld project contains a number of namespaces:

  • AIWorld: Contains the entry point of the application.
  • AIWorld.Core: Contains a number of datastructures used troughout the project.
  • AIWorld.Drawable: Contains various 2D and 3D drawable compoments.
  • AIWorld.Entities: Contains the entity interfaces, base types, and a number of entity implementations.
  • AIWorld.Events: Contains event arguments objects.
  • AIWorld.Fuzzy.x: Contains all classes required for computing fuzzy logic.
  • AIWorld.Goals: Contains the classes used for the implementation of goal driven behavior.
  • AIWorld.Helpers: Contains a number of static helper classes.
  • AIWorld.Planes: Contains classes of drawable planes.
  • AIWorld.Scripting: Contains classes and interfaces related to scripting.
  • AIWorld.Services: Contains every service available in the application.
  • AIWorld.Steering: Contains every steering behavior available in the simulation.


There are three classes which implement the IEntity interface. The Agent, Projectile and WorldObject class.


Agents are moving entities with a behavior specified by a script. The script can define the behavior of the agent by manipulating a couple of data containers within the agent, including a list of active steering behaviors, a stack of path nodes (See Path Planning chapter), a stack of goals and a dictionary of model mesh data.


Agents can fire a projectile from a specified position with a specified velocity. When the projectile hits an entity which implements the IHitable interface it will call the IHitable.Hit method on the hittee. If this method returns true the projectile will remove itself from the world. When spawning a projectile there is also an option to specify a number of seconds after which the projectile is automatically deleted to prevent the world from filling up with thousands of projectiles.


World objects can be placed by the main script as obstacles for agents and projectiles. When creating a world object there is an option to specify which meshes of the model should be visible and to supply a rotation, scale and translation matrix.


AIWorld requires scripts to be written in the Pawn Language. A language guide can be found in docs/Pawn_Language_Guide.pdf.

Every instance of the pawn abstract machine is created by the ScriptBox class. By default only a small number of core native functions (functions defined outside of the abstract machine) are available to the script. The class contains three methods to make more native functions available to the script. Using ScriptBox.Register(string, AMXNativeFunction) you can register a function to the script with the specified function name. ScriptBox.Register(AMXNativeFunction) has the same behavior, but uses the function name of the specified function itself. ScriptBox.Register(object[]) registers every method tagged with the ScriptingNativeAttribute in the specified object instances.

In the engine, there are three types of scripts. Each of these scripts have access to some specially crafted native functions to allow the script to fulfil its purpose. Some services (see the Services chapter) also provide functionality to specific scripts.

Main Script

The main script creates the terrain, objects, agents, graphs and global variables. It can also handle user input, such as mouse clicks and key presses. The main script instance is created by the main Simulation class. To this script only a limited number of native functions are available. These functions can be found in the bin/include/ file. Functions provided by the GameWorldService, ConsoleService, DrawingService and SoundService are also available in this script.

Agent Script

The agent script defines the behavior of an agent. The agent script instance is created by an instance of the Agent class. To this script a large number of native functions are available. These functions can be used to set properties, manage the active steering behaviors, manage the goals, change the position and add path nodes to the path stack (see the Path Planning chapter) of the agent. There are also a number of native functions available for calculating the distance to a point and transforming vectors between local and world spaces. These functions can be found in the bin/include/ file. Functions provided by the ConsoleService, DrawingService, GameWorldService, SoundService and FuzzyModule (see Fuzzy Logic chapter) are also available in this script.

Goal Script

The goal script defines the behavior of an agent while the goal is active. The goal script instance is created by an instance of the Goal class. To this script a couple of unique native functions are available. These functions can be used to manage the subgoals of the goal. These functions can be found in the bin/include/ file. All agent script native functions are also available to a goal script. Functions provided by the ConsoleService, DrawingService, GameWorldService, SoundService and FuzzyModule (see Fuzzy Logic chapter) are also available in this script.



The camera service contains various methods to allow users to smoothly operate the camera. When the camera is manipulated, it automatically updates a view matrix. The service also watches the graphics device for changes and updates the projection accordingly.


The console services contains methods to append messages to the console or to the chat. Using the Tab key the console can be toggled on and off. The chat, however, is always visible. The chat can only hold a limited number of messages which disappear after a few seconds. The console service provides a number of native functions for scripts to append messages to the console or chat.


The drawing services provides native functions for scripts to draw shapes or other components in the simulation. Available components include a line, cone, cylinder, sphere, 2D (overlay) text and 3D (label) text. When the script creates a component, a handle to the component is returned. This handle can be used to update any property of the component using the functions provided by the service. The drawing logic of the components can be found in the AIWorld.Drawable namespace.


The game world service provides methods for finding entities in and adding entities to the game world. It also provides native functions to scripts for creating graphs and accessing them, manipulating properties and variables of world objects and agents, manipulating global variables shared across all running scripts and sending messages or calls to other scripts.


The particle service has sadly not been implemented yet. It is meant to allow scripts to spawn particle effects in the game world.


The sound service provides methods and native functions for scripts to play a sound effect at a position within the game world.


The Agent class contains a list of active weighted steering behaviors. The agent script, its goals or its goals' subgoals can add or remove these steering behaviors. Any class within the assembly which implements the ISteeringBehavior interface will be accessible from the scripts. For every of these classes, a function called Add[classname] will be registered to the script, where [classname] is the name of the class of the steering behavior. If the class name ends with SteeringBehavior, it will be omitted from the function name.

The different steering behaviors are combined by multiplying the calculated force by the weight, summing these values and then truncating it to the maximum force for the agent.


The alignment steering behavior attempts to align the heading of the agent into the same direction as nearby agents. The script can specify which agents to take into account by specifying a variable name for the agents to have and the value it should contain.


The arrive steering behavior attempts to steer the agent towards a specified target and stopping it at the target position.


The avoid obstacles steering behavior attempts to avoid the agent hitting solid entities.


The cohesion steering behavior attempts to let the agent stay close to nearby agents. The script can specify which agents to take into account by specifying a variable name for the agents to have and the value it should contain.


The evade steering behavior attempts to keep the agent away from agents within a specified range. The script can specify which agents to take into account by specifying a variable name for the agents to have and the value it should contain.


The explore steering behavior attempts to zigzag the agent across the specified area.


The flee steering behavior attempts to move the agent away from a specified point.


The offset pursuit steering behavior attempts to tail the specified agent at the specified offset.


The pursuit steering behavior attempts to tail the specified agent.


The seek steering behavior attempts to go straight towards the specified point at top speed.


The separation steering behavior attempts to keep some distance from nearby agents. The script can specify which agents to take into account by specifying a variable name for the agents to have and the value it should contain.


The stop steering behavior attempts to stop the agent with the specified breaking weight.


The wander steering behavior attempts to let the agent wander in a random, natural way.

Path Planning

When an agent generates a path, the resulting path nodes are pushed to the path stack in reverse. This way, the script can keep popping nodes off and seeking towards it until the stack is empty.

Graphs are filled using a self invented algorithm. The algorithm is displayed in the following pseudo code:

FillGraph(name, minX, minY, maxX, maxY, offset)
    offsets = {
        Vector2(offset, offset),
        Vector2(-offset, -offset),
        Vector2(offset, -offset),
        Vector2(-offset, offset),
        Vector2(-offset, 0),
        Vector2(offset, 0),
        Vector2(0, -offset),
        Vector2(0, offset)

    for every x between minX and maxX in steps of offset
        for every y between minY and maxY in steps of offset
            point = Vector2(x, y)

            entities = solid entities near point

            for p in point + each offsets
                if a ray cast from point to p hits no entry in entities
                    add point to graph

The A* algorithm implemented in the engine works according to the following pseudo code:

ShortestPath(start, finish)
    nodes = list of Vector2
    result = list of Vector2

    add node from graph to nodes where node position is start

    set every node's distance in graph to infinity

    while nodes is not empty
        current = node from nodes with lowest (node distance +
            (manhattan distance between node and finish))

        remove current from nodes

        if current position is finish
            return the result

        for each edge in current node
            if current distance + edge distance < edge target distance
                edge target distance = current distance + edge distance
                edge target previous node = current

                add edge target to nodes

The algorithm has, for example, built the following path:

The blue line is the path the algorithm has generated.

I meant to display all evaluated nodes using red lines, but due to a lack of time it currently only shows the nodes whose previous node is within the path.

Fuzzy Logic

I've implemented the fuzzy logic classes mostly in the way the book Programming Game AI by Example describes it. I have however moved the data container of fuzzy rules to the consequence variables of the rule. Because I wanted to create and use the fuzzy logic from the script, I've written a simple fuzzy logic interpreter which can be found inside the FuzzyModule class.

Temple of Kotmogu


Temple of Kotmogu is a gametype(battleground) in the popular game World of Warcraft. The rules of the gamemode has been recreated within this simulation. Two teams of tanks battle to collect the yellow orbs. Every few seconds your team earns a few points per orb in its possession. The closer the carrier of the orb is to the center of the map(the center is the most powerful area), the more points it receives per every few seconds. Every tank has a limited number of ammo and health. The tanks will attempt to shoot down the other team. If a tank dies and it carried an orb, the orb will be returned to its point of origin. A few seconds after a tank has been shot down, it will reappear at its spawn point.

Every few seconds a plane arrives with a carepackage which is dropped at a random position on the map. The carepackage contains a random amount of ammo or a repair kit for a tank to increase its health.

The tanks have a number of goals:

  • attack: search and destroy the nearest enemy.
  • combat: fight tanks within firing range.
  • defend: defend the nearest ally with an orb.
  • get_carepackage: get the nearest carepackage.
  • get_orb: get the nearest orb.
  • hold: hold the orb and don't leave the power area.
  • powerup: find a more powerful area.
  • think: think which goal to work on next.

All goal scripts are heavily commented and can be found in bin/kotmogu/tank/goal.


By clicking on a tank you can selecting it. Selected tanks display their debug information (including the active path) on the screen.

By right clicking while holding the left shift while you have selected a tank, the tank will calculate a path to the clicked point and will navigate to there.

By holding the F9 key all graphs will be displayed on the screen.

Fuzzy Logic

In the simulation, there are currently 14 fuzzy variables:

  • ammo: The amount of ammo the tank has.
  • area_power: How powerful the area the tank is in is.
  • carepackage: The distance to the nearest carepackage.
  • enemy_orbs: The number of orbs the enemies have.
  • have_orb: Whether the tank has the orb.
  • health: The health of the orb.
  • orb: The distance to the nearest orb.
  • orbs: The number of orbs the allies have.
  • attack_desirability: Desirability to attack.
  • carepackage_desirability: Desirability of a carepackage.
  • defend_desirability: Desirability to defend.
  • hold_desirability: Desirability to hold this position.
  • orb_desirability: Desirability of an orb.
  • powerup_desirability: Desirability to get to a more powerful area.

Because of the big number of FLV's and the short amount of time I have to finish this project, I wont add the graphs for every FLV to this document. The graphs are created and can be found in bin/kotmogu/tank/common/ at line 28.

In the think goal of the tanks I use these fuzzy variables to decide which action to take:

public OnUpdate(Float:elapsed)
    // ...

    // Fill the fuzzy variables (e.g. health, ammo, etc.)

    // These functions are defined in bin/kotmogu/tank/common/
    new Float:want_orb = GetOrbDesirability();
    new Float:want_powerup = GetPowerupDesirability();
    new Float:want_carepackage = GetCarepackageDesirability();
    new Float:want_defend = GetDefendDesirability();
    new Float:want_attack = GetAttackDesirability();
    new Float:want_hold = GetHoldDesirability();

    // Macro which returns true if the specified desirability is the highest of
    // all desirabilities
    #define BEST_DESIRABILITY(%1) (%1 >= want_powerup && \
        %1 >= want_carepackage && %1 >= want_defend && \
        %1 >= want_attack && %1 >= want_hold)

    else if(BEST_DESIRABILITY(want_powerup))
    else if(BEST_DESIRABILITY(want_carepackage))
    else if(BEST_DESIRABILITY(want_hold))
    else if(BEST_DESIRABILITY(want_defend))
    else if(BEST_DESIRABILITY(want_attack))

    // ...


Due to my idea of building a whole simulation engine instead of just this one simulation, I've had to spend a lot of time working on this project. Because it was such a big project and the time was limited there are some things I've not had time for to implement, improve or fix.

I have, however, had lots of fun working on this project and I have learned loads of stuff about artificial intelligence in games. I've no encountered any big problems in this project. The biggest challenge has been to write a wrapper for the pawn abstract machine which in the end works perfectly and very fast.

Room for Improvements

A number of the following items are also annotated in the source code.

  • After you navigate a tank to a position by shift-right clicking in the simulation, the stack of goals in the debug overlay looks weird. I've not had time to look into this.
  • Tanks shoot at the nearest enemy, even if there is a friendly tank or a tree in between them.
  • Cohesion, alignment and/or separation can cause tanks to move in odd ways if nearby tanks are standing still or if there are no nearby tanks.
  • Not all of the official Temple of Kotmogu rules have been implemented, including "Delivering a killing blow to one of your enemies will net your team a significant number of victory points." and "Orb-holders deal and receive increased damage, and the healing they receive is reduced. These effects are minor at first, but they increase quickly over time while the orb-holder lives.".
  • I've started of documenting every method and function, but do to time pressure not all methods and functions have proper documentation anymore, both in the script and the engine.
  • The powerup goal tries to reach the other side of the map to power up. This normally results in the tank driving into a area with a higher power anyways, but occasionally it does not, which can result in the tank being stuck in the powerup goal.
  • Sometimes tanks can get stuck behind an obstacle for a while.
  • Accessing an entity by id in the GameWorldService makes the service loop trough all the entities until it has found the entity with the specified id. This has a speed of O(n) It would be much better if the entities are put in an AIWorld.Core.Pool<T> as well, which can get the associated object of an id in O(1) time.
  • Every time a model, sound or texture is loaded, it's loaded from Simulation.Content again. I expect this class to read the object from the file after every call. It would be much better if the engine keeps these instances in memory. The same goes for scripts. Scripts are loaded very often and are loaded from the disk every time.
  • The fuzzy logic used in the simulation could use lots of improvements. I have not spend a lot of time on designing the FLVs and the rules.


The source code and guides of the Pawn Language can be found on:

The rules of the Temple of Kotmogu can be found on:

Various models and sounds were derived from:

During research I've used the code of Programming Game AI by Example which can be found on:


A 3D simulation engine I've written for a school assignment.



You can’t perform that action at this time.