Goal Oriented Behavior
A Technique For Complex AI Behavior In Games
Neal Wojtowicz
Goal Oriented Behavior is an artificial intelligence technique that allows AI agents to make intelligent decisions and generate more complex action sequences. It adapts the Stanford Research Institute Problem Solver (STRIPS) which was an automated planning system created in 1971 and implemented into one of the first intelligent robots, SHAKEY (Cassel 2017). Goal Oriented Behavior was first pioneered by the game F.E.A.R for creating "cinematic" combat situations and continues to appear in games today.
What is
Goal Oriented Behavior?
Goal Oriented Behavior provides autonomous and intelligent agents through goals and actions. A goal is a state that the agent wants to satisfy, like attacking an enemy or moving to a location. Essentially, goals are what the agents want to accomplish and each goal has a list of actions that the agent can execute. The goals and actions chosen are based on their insistence values.
Insistence Values
How do agents choose goals and actions? The answer is through something called insistence values. Each goal and action has a calculated priority value, the insistence value, that measures their importance to the agent based on preconditions (AI and Games).
For example, say the agent has a goal to attack an enemy. Does the agent have a weapon? Does the agent need a weapon? Is the target in range? These are questions that Goal Oriented Behavior answers through these insistence values. If the agent does not have a weapon, the system will recognize it and adjust the insistence values so that retrieving a weapon has more priority than attacking without one, thus simulating the intelligent behavior (AI and Games). This also means that each enemy agent in the scene will have different behaviors because they are calculating different insistence values (Owens 2014).
Goal Oriented Behavior
vs
Finite State Machines
Goal Oriented Behavior at its core is driven by a Finite State Machine (FSM), although its architecture is still different than a complex FSM. Complex FSM's can be real headache since all actions are connected to each other, seen here below (Chaudhari 2017).
These connections (or really transitions) can cause problems when debugging, maintaining, or changing actions. If you wanted to remove an action, you need to update the other actions that are affected by it. Removing the action "get log" would be a real pain to clean up. Goal Oriented Behavior avoids this by acting as a manager for actions, effectively decoupling the actions from each other. Only the goal system and goals manage when these actions are called. The diagram below showcases the difference between an FSM and GOB architecture.
Applications in Games
F.E.A.R
The game F.E.A.R first pioneered this idea with its behavior system GOAP or Goal Oriented Action Planning. The developers wanted to create intense combat sequences, but complex FSMs made this difficult. Instead, they adapted STRIPS towards game artificial intelligence and created a great example of Goal Oriented Behavior (AI and Games).
The developers implemented a system that had a total of three game states or goals, Goto, Animate, and Use Smart Object (AI and Games). Goto would handle movement, Animate handles animations from generated actions, and Use Smart Object is just interacting with game objects. Enemies had over 120 actions and around 80 of them were for combat (AI and Games). They also calculated priority values for those actions, simulating the intelligent behavior, and created action lists to execute complex behaviors. The game was praised for its extremely fun combat situations.
Most games today use behavior trees or Goal Oriented Behavior when implementing complex AI behavior. Action games, like F.E.A.R, very much benefited from the ease of maintaining and iterating on enemy agents' actions. The complex FSM was not holding the developers back and they created some very nice sequences from implementing this technique.
Tech Demo
I created my own tech demo showcasing Goal Oriented Behavior and its insistence values in action. Essentially, the red block represents an enemy agent who has two goals, ATTACK and RELOAD, and a green block that represents the target. Actions include RELOAD_GUN, OPEN_FIRE, and CHASE. You can move the target with WASD to simulate chasing. The ATTACK goal simply generates attack actions for the target. Pressing TAB also spawns more enemies in the center. This is a very basic example of this technique but showcases the core idea of Goal Oriented Behavior, intelligent and dynamic enemies.
Link to in-browser tech demo: https://nealwojtowicz.itch.io/goal-oriented-behavior
Conclusion
In conclusion, Goal Oriented Behavior is an AI technique for games that is designed to provide intelligence for agents. Unlike complex FSM's, you can easily change, create, and remove actions without having to change other actions. This allows for you to simply plug actions into a goal, as well calculating the insistence value for it, and improve your game agent behavior. Overall, Goal Oriented Behavior is a great game AI technique and can generate complex AI behaviors.
Sources
-
“Building the AI of F.E.A.R. with Goal Oriented Action Planning, AI and Games.” AI and Games, 6 May 2020, www.aiandgames.com/2020/05/06/ai-101-goap-fear/.
-
Chaudhari, Vedant. “Goal Oriented Action Planning.” Medium, Medium, 12 Dec. 2017, https://medium.com/@vedantchaudhari/goal-oriented-action-planning-34035ed40d0b
-
Owens, Brent. “Goal Oriented Action Planning for a Smarter AI.” Game Development Envato Tuts , Envato Tuts, 23 Apr. 2014, https://gamedevelopment.tutsplus.com/tutorials/goal-oriented-action-planning-for-a-smarter-ai--cms-20793
-
Cassel, David. “Remembering Shakey, the First Intelligent Robot.” The New Stack, 6 Mar. 2017, https://thenewstack.io/remembering-shakey-first-intelligent-robot/