F.E.A.R. Designer Diary #2 - A Study of Smart AI, Part II

Gamers are still buzzing about the incredible artificial intelligence demonstrated by the enemy soldiers in the recent demo of F.E.A.R., the upcoming first-person shooter that's also one of the most anticipated games of the year. By now you've probably come to expect that the enemies in most first-person shooters are there to serve as cannon fodder. However, the elite soldiers in F.E.A.R. actually play like human beings who want to win and survive. Soldiers will literally jump over barriers, flip tables over for cover, crawl beneath obstacles to get to you, and scream commands to one another as they hunt you down. It's a pretty amazing technical achievement. Thankfully, the team at Monolith was kind enough to give us a glimpse into how the AI in F.E.A.R. is so lifelike.

You'd better bring your "A" game, because these guys are tough.

F.E.A.R. the AI, Part II

The goal of artificial intelligence in video games is to create opponents that challenge you with realistic competence and provide you with a rich and varied game experience. Recent games have employed scripted behaviors with a certain degree of success, but a careful scrutiny of enemy behavior quickly exposes the limitations of that technology. Scripted behavior leads to repetitive and often inappropriate reactions that break the immersive feel of a game. As the game experience of F.E.A.R. is based on confronting a highly intelligent, highly trained combat force that quickly adapts its tactics in response to the player's actions, a new AI paradigm had to be created. Fortunately, the F.E.A.R. development team possessed the unique combination of proven experience and a thirst for applying the latest academic advances in AI research.

The results of the team's hard work are some of the most startlingly intelligent enemies you have ever seen. Gone are the days of replaying a segment of a level with the confidence of knowing your enemies' paths and reactions. In F.E.A.R., your enemies formulate real-time plans based on environmental circumstances and your actions within that environment. Your opponents will move and react to you with the familiarity and tactical awareness of a seasoned multiplayer veteran. You will witness AI opponents change their courses to avoid danger, manipulate the environment to create cover, flank you, and dynamically support squadmates with cover fire and verbal warnings. Your opponents in F.E.A.R. work together in such a realistic way that you may compare the single-player combat to the tactics and strategy of a quality multiplayer match.

Recent advances in AI have allowed game designers to give their non-player characters basic decision-making capabilities with predefined finite state machines (aka FSM). This standard FSM technology allows for goal-directed characters that are able to activate a goal in response to a situation. However, as Jeff Orkin, the game's AI architect, notes, "Once a goal is activated, the character runs through a predetermined sequence of steps, hard-coded into the goal. The embedded plan can contain conditional branches, but these branches are predetermined at the time the goal is authored." While this can sometimes give the impression of a "thinking" NPC, this is essentially a mildly more versatile form of scripting. The characters have limited and predetermined behavior patterns, and the rigidity of the "sequence of steps" often forces recognizable "dumb" behavior that breaks the immersion of the game.

These soldiers actually have goals...like killing you.

The AI takes the next leap in the evolution of AI decision making with goal-oriented action planning (aka GOAP). In addition to deciding to activate goals, the AI also plans how to accomplish these goals in real time. "A character that formulates his own plan to satisfy his goals exhibits less repetitive, predictable behavior and can adapt his actions to custom-fit his current situation," says Jeff. Goals in GOAP are not created with a hard-coded plan. Instead, GOAP simply defines the conditions necessary to satisfy a goal, and the character determines the steps to satisfy this goal in real time. With this structure, the AI is able to dynamically replan to react to environmental factors. If a situation changes, the NPC recognizes this because the steps planned to accomplish his goal are no longer valid. When a plan becomes invalid, the NPC reassesses the situation and either finds alternate means for accomplishing the goal or activates a different goal. For example, if the player slams the door on a pursuing NPC, the NPC replans and decides to kick the door in. If the player physically blocks the door with his body, the NPC replans again and decides to dive through the window. Creating individuals that act intelligently is only half the battle. The next step was to make them act as an intelligent team.