Medal of Honor: Underground
|Medal of Honor: Underground|
|AI Era:||Tactical Reasoning|
Medal of Honor: Underground is a videogame that is the prequel to the World War II hit Medal of Honor. In Underground, the player takes the role of Manon Batiste (Based on Helene Deschamps Adams), a French woman who appeared in the first game as an adviser.
Description of AI Behavior
The AI in the original MoH game had been fairly rich. The enemies would use cover, attack in groups, and even throw back grenades the player tossed. In Underground, we wanted to further enrich that set of behaviors, as well as add some unique flavor to the game. Because of the incremental nature of this game, the AI had a huge amount of triggers that the AI would respond to. The inclusion of Disguise Mode (where the main character would dress as a journalist, and infiltrate areas in a non-combative way) meant there was also a large number of non combat behaviors that needed to be built for the AI characters as well.
One of the things that the original game didn't have was any vehicles, from either side of the joystick. You were always on foot, and were always fighting against other soldiers on foot. The first thing I worked on for this game was AI controlled tanks, which added a lot of WWII flavor to the game. Also added the motorcycle level, which was a major departure in gameplay from the norm (it was more of an on rails shooter than the open world FPS game) and so it worked very well. Also, since the original game already had a nice complement of weapons, we were able to add some of the more esoteric weapons into the game (as well as having the AI characters use them). The AI had to be given behaviors to use and deal with all these new gameplay elements. The game also had a ton more story specific behaviors because the backstory was actually a much larger part of this game than in the first Medal of Honor. Finally, the game had a full "companion" that was fully AI controlled and would either follow you through the level and help, or in some cases go off on his own and meet you in later parts of the level. This was a very large undertaking, since we didn't want this companion to be a sore spot if he acted stupid or didn't help in appreciable ways.
Scripted trigger based State Machine
The entire game was created using small event driven scripts. These scripts could turn on/off various event triggers, set variables, play sounds, assign AI animations, or fire other events. Any individual event script was quite simple. The magic behind this game was the sheer amount of event scripts that comprised the AI. There were hundreds of reactive scripts that could be available during open gameplay, and hundreds more that were more serial in nature for story elements. For the most part, behavior collisions were handled with a very simple number based priority system, as well as randomly choosing when multiple behaviors of the same priority were available. Again, the architecture wasn't the key here. It was the sheer amount of AI content involved.
|This section is based on|
by AIGPG member,
[Your Name Brian Schwab].
Because of the richness of the total behavior set, the game really felt immersive and alive. The enemies had an enormous amount of potential behaviors at any given time, and as such didn't seem repetitive or overly scripted, instead there were many small atomic actions blending together into a much more organic sort of AI.
Many different types of interactions
The first Medal of Honor game was all combat, just in different locales. In Underground, we wanted to enrich the experience by giving the player other things to do. Some missions you were heavily involved with your AI companion. Other missions, you were disguised as a journalist and taking pictures of guards. There was a much larger bank of types of interactivity with Underground then there was before.
What Didn't Work
Typical State Machine Issues
Classic state machine problems. Eventually, you get lots of reactions to a particular event, which means decision collisions. The trick is to just allow them, and randomly choose amongst the collisions. Making a ton of "equivalent" reactions that would be chosen between made the game seem much less deterministic and rich. Making very specific behaviors does nothing but make new categories of behavior that you then have a limited number of reactions to. This game required a lot more story moments than the first MoH, so there were a lot more specific event reactions. These definitely didn't feel as rich as the rest of the game, but this really only affected replay value, as opposed to how fun it was to play.
Because the behaviors were described as a series of small scripts, the overall picture of behavior connectivity was largely in the author's head. There was no tool for visualizing this connected web of behaviors, and as such following the thread of potential AI activity could be tedious or even difficult to follow. Many greps died to bring you this information.
Reception by Public
Pretty successful overall. Game was reviewed well, and seen as a worthy sequel. There were still issues, but they were more because this was later in the life cycle of the PS1, and as such the next generation of game systems had started to come out (Sega Dreamcast, PS2 had just come out). So the lower graphic fidelity, much lower polygon counts, and some of the classic fixed point collision issues were not seen in the same light as they would have been just a year earlier.