Zubair Chaudary, Kenny Rentschler, Constantin Savtchenko Game Genre that involves projectile weapons Played “though the eyes” of game character First Person point of view Creates a game play, AI, and game design Issues Popular FPS Titles Quake, Doom, Wolfenstein 3D, Goldeneye 007, Half Life, Halo, Call of Duty, Counter Strike, Unreal Tournament… First known FPS: Maze War (1974) FPS Stepping Stone: Wolfenstein 3D (1992) Intro of Deathmatch: Doom (1993) 3D Polygons: Quake (1996) Major Console FPS: Goldeneye 007 (1997) Entirely Script Based FPS: Half-Life (1998) Exceptional Plot and Story: Half-Life (1998) Team play focused FPS: Unreal Tournament (1999) New Millennium FPS: Halo (2001) MMOFPS: Planetside (2003) FPS as Story Medium: Bioshock (2007) Graphics Killer: Crysis (2007) $1 Billion Dollar FPS: Call of Duty – Modern Warfare 2 (2009) Arguably the most valuable to the gaming industry Envelope Pushing sector Graphics, Game Play, AI, Audio Genre Bending Has created spinoff genres FPS medium allows developers to incorporate elements from other genres Movement Layer Does not determine where to move to, only how to do so. Avoid obstacles, follow characters, and path finding through complex environments Animation layer Selecting, parameterizing, and playing character animation sequences Generating specific animations that are situational, and not prerendered/coded Combat Layer Assessing characters current tactical situation Selecting tactics in combat Performance is key, as this is the fundamental aspect of FPS Behavior Layer Determines character’s current goal, and attempts to reach its goal Pathfinding Dynamic Obstacles Spatial Reasoning Algorithms and methods to find most efficient movement from point A to Z Most FPS levels tend to be static, so AI use a pre-generated Database Allows for A* search algorithms Situation changes significantly with the addition of dynamic level elements Player can move objects in front of AI paths Level Destruction can block paths If the AI does not adapt, will walk into a wall and get stuck. Handling of dynamic objects Some games ignore them (pseudo random movement) Introduction of Local Pathfinding Uses Global Pathfinding to establish a general path from A to Z Create lots of “waypoints” in-between A-Z Local Pathfinding determines if subpath A-B are clear If they are, continues to waypoint B, where it evaluates B-C If not, it computes an alternate path to waypoint B Local Pathfinding also uses A* search AI Data generation tool 3D Pathfinding Part of the Unreal 3 Engine Management of perceptions (sight, smell, and sound) Subsystem that performs tasks assigned by higher level systems Performs discrete movement commands ▪ Move to (X,Y,Z) Only be executing one movement command at a time. Movement controller is an object that owns the current movement command Can be use to handle different kinds of movement, like walking or swimming When AI enters Combat, almost all control is given to combat controller Attempts to assess tactical situation Number of enemies, direction, health, ammo More Advanced: cover locations, camping locations Issues commands to other systems Movement, aiming, shooting Most difficult task for AI combat controller Spatial Reasoning requires understanding of level (Spatial Configuration) Needs to know where things are, and how objects are going be use 3-Dimensonal Levels increase complexity significantly Pathfinding solution can be used Pre-generate database of objects and their significance Works poorly with dynamic levels Level Designers can embed ‘hints’ Makes Level design cumbersome Error prone Customized tools analyze geometry Can be run dynamically to populate database Physical properties of objects? We will use Valve’s Left 4 Dead to discuss and examine spatial reasoning. Things to think about: Navigating In 3D Optimal Paths CPU Calculations Example comes from a presentation created by Michael Booth. Path finding algorithms help AI’s navigate optimally through environments. How do you define an environment? By using a Navigation Mesh!! • Used instead of waypoints • AI can move anywhere in between the grids. However note the Jagged Path. We could hardcode a “smooth” path… Or we could use what Valve calls “Reactive Path Following”… Look ahead in the path and calculate Use local obstacle avoidance Oh, it’s a crate!? We would press spacebar, and bunny hop over it, should be easy for computers right? Wrong AI has to deal with the environment as we do: Quickly Efficiently Humanly Combining the two strategies we examined allows for an AI that deals with the presented necessities An Example of a Tactical Situation Current Location ▪ In cover, or out in the Open? Ammo (unless they cheat) Enemy ▪ How Many? ▪ From where? Current Objective(s) Behavioral Situation Tactic Selection Problem Nature of Tactic under Consideration Relative Tactical situation of all combatants Current Tactical Situation Use of Tactical Library is common FSMs Databases Databases with Q-Learning Basic Tactics Camp ▪ AI hides in a location, and waits for an opponent to appear ▪ Effective, but behavior can seem very scripted Joust ▪ AI rushes opponent, while firing ▪ Runs past opponent with the hopes of being able to turn around and rush from behind Circle of Death ▪ AI circles opponent while firing ▪ Keeps a variable radius away from target Ambush ▪ AI ducks behind cover, periodically pops out of cover and shoots at enemy. ▪ Similar to camp, but AI must know where enemy is coming from Flee and Ambush ▪ AI runs away from enemy, and then takes cover and gets ready to ambush As games evolve, game designers strive to give players more freedoms: Dynamic level designs Interactive objects Large, multi-approachable situations We will study how AI has kept up through Killzone©’s implementation. Old AI would use “hints” and series of “ifthen” scripts. Hints remain static, rely on game designers Enormous amount of effort for responsive AI’s to react to situations and terrain The AI must take into account: Multiple approach points Dynamic environments Dynamic positioning End Result: Responsive AI AI analyzes the situation Inputs are dynamic Using gathered information, the AI acts: Goal setting/discarding On-fly-algorithms allow for on-the-fly decision making Tactics are procedures, not scripts Summary: tactics are _______ based on the ___________ and the _________. Killzone AI Summary Accomplish goals, or change based on desirability/feasability Move To Position, Do Action Killzone’s describes situations: Cover from threats Lines-of-fire Danger zones Area of operation Decisions made using: Position evaluation Line-of-sight and Line-of-Fire The assignment of values to the waypoints of Killzone Some factors that give higher values Distance to primary target Amount of cover Not in the way of friendly-fire Wall hugging Outside of danger-zone (grenades, tank shells) Position Evaluation Helps Movement Tactical Positioning Tactical Movement Position evaluation functions assign values to the waypoints of Killzone Tactical Movement is smarter movement Movement usually done by “lowest weight” Notice we can use some of the existing position evaluation algorithm to make “dumb” paths cost more ▪ Line of fire ▪ Close proximity to danger zones ▪ Running at the target Position evaluation algorithms, can also be used for attacking Indirect Fire ▪ Different determiners can be used to predict where to throw grenades Suppression Fire ▪ Other determiner can be used to figure out at which cover to lay suppressive fire onto Advantages Parameterization – change values, not code Easily adapted for other uses, such as multiplayer Procedural tactics are easier to adjust, and work with – compared to scripts/hints Removes extra ray-casts Disadvantages Difficult to debug, the AI is “on its own” – compared to scripts/hints Also, difficult to get it to do your bidding – if you want to actually script something Opponent Selection How to Target? ▪ Most FPSs pit all AI against an enemy ▪ Kung-Fu tactics? Need for Targeting Heuristic ▪ Tactical Situation Awareness ▪ Worry about defending itself first ▪ Most vulnerable, nearest target ▪ Ranking function can simplify Need to be able to change targets based on changes in Tactical situation To-Hit Roll Less than perfect aiming for believability Calculating factors ([0.0,1.0] range) ▪ AI skill ▪ Range ▪ Size ▪ Relative target velocity ▪ Visibility ▪ Target state ▪ AI state Probability to hit = skill*range*size*…*state Aiming Cheat Angle ▪ Maximum that a projectile can differ from gun barrel while still being believable ▪ Doesn’t need to match line of sight Location ▪ Point with high probability of being hit ▪ Model target as overlapping boxes and pick center of volume Shooting and Missing Pick point close to target ▪ Low probability of hitting Should be visible to player Missing when close and hitting from a far distance seems unreal Ray Testing Test projectiles trajectory before firing Avoid obstacles and friendly fire Dead Reckoning Estimation of enemy position at given time Shoot where the enemy will be when bullet arrives Weapon Trajectories Difficult to predict slow moving projectiles (arrows) Use physics in ray testing to test trajectory Collision Notifications Projectiles store pointers to object that fired it ▪ On hit or miss object is notified Helps firing object adjust targeting Radius Testing Helpful for area effect weapons Treats everyone as cylinders or spheres Tests area damage on everyone to avoid friendly fire and maximize enemy damage Responsible for controlling current state and high level goals Most FPS use a Finite State Machines Idle – Standing Guard Patrolling – AI following a designed patrol path Combat – AI is engaged in combat and most control is given to combat controller Searching – AI is looking for an opponent to fight or searching for an opponent who fled Summoning Help – AI is searching for help sound, patrol -> ɛ quiet, ɛ -> patrol search patrol kill, ɛ -> patrol kill, ɛ -> search see, search -> ɛ see, patrol -> ɛ fight Scripting and Trigger Systems Need scripting and triggering events to tell the AI to do something Triggered events set AI parameters Send commands to various subsystems ▪ Changes state from Idle to Searching, or Searching to Combat Disadvantages? Adds fairness to the game Even though the game knows where a player is, the AI needs to “see” or “hear” a player approaching Realism A sleeping enemy has to be startled to enter combat mode Sneaking element is possible and quantized Visual Use ray casting to look for players Distance, angle, and visibility factor into vision Auditory AI needs to receive sound notifications Everyone in earshot should hear the sound Priority of sounds (e.g. bird chirping and gunshot) Audio Occlusion ▪ Interference of sounds to create noise Tactile(touch) Alert when run into or wounded AI is initially designed to be very smart (usually) Based on designers, the AI is given flaws to dumb it down Also changed by difficulty settings ▪ Affects parameters for targeting, etc Player Modeling ▪ On the fly difficulty adjustment Aiming very similar Movement may change How to move so the AI does not block a teammate’s vision? Behavior How to choose objectives?