Adjustable Group Behavior of Agents in Action-d Games Westphal, Keith and Mclaughlan, Brian Kwestp2@uafortsmith.edu, brian.mclaughlan@uafs.edu Department of Computer and Information Sciences University of Arkansas - Fort Smith Abstract Non-player characters (NPCs) within a single-player action-d game have historically been easy to kill and often do not work as a team to overwhelm the players. In addition, they generally do not adjust strategies d on the player s behavior. This research proposes that these s should have adjustable behavior parameters that allow them to adapt to the behaviors of their opponents. In particular, this research proposes these NPC s have an adjustable ness towards the player giving them an option on whether to hide or to attack. Also, it examines the utility of s that attempt to attack together and work in a group using simple decision making rather than being individuals. We developed a simulation that gathers information regarding the win percentages of teams of s that have various behavior combinations versus different types of opponents. Behavior parameters examined include individuality versus teamwork and posture versus defensive posture. Results showed that certain behavior combinations have significant utility when the behavior of the opposing team is identified. Index Terms Artificial Intelligence, Agent, NPC, Group decision making, MMO (Massively Multiplayer Online Games), and game loop I. Introduction AI-controlled non-player characters (NPCs) are the basis for opposition in single-player and cooperative team video games. These NPCs have historically been easy to destroy and often do not work as a team to overwhelm the players. In addition, they generally do not adjust strategies d on player behavior. When a skilled player plays against NPCs, the game can become rather easy. Fortunately, computer opponents have evolved from the simple behaviors found in games such as Pac Man and Pong into the realistic war games of today such as Killzone, Modern Warfare, and the Battlefield games. With these recent increases in AI technology in video games, the AI of a game has become a significant selling point. Modern game reviewers now keep a place on their notepad just for rating the AI in these games, and when many players hear that the game has a great AI, they are more likely to want to play it. Unfortunately, most of a game s AI resources are focused on individual NPC tactics [1]. In the latest generation of video games, the AI controlling the NPCs will hide if shot at, will also run away from grenades, and will look for advantages locations. However, they are far from acting as a team. Some of the best games out do have communication between their AI s to assist in making the game more difficult for the player. Nevertheless, communication alone is still rather far from group decision-making. In action-d games such as FPS games, going solo is almost guaranteed to get you killed being as the best you can hope for is a 1-on-1 match up. However when running in groups, s have a much better chance to come across match ups that are more in their favor. This paper postulates that coordinating the group behavior and adjusting it d upon the behaviors of the opposing force can lead to increased success for the NPC team and a more rewarding play experience for the players. Currently, the primary work being done in group decision-making in the field of Artificial Intelligence is in the game of football (soccer in USA) [2]. Being a complex sport with many rules, this research is a great starting point within the field of group decision-making. However in the case of action games where destruction is very possible, it is not the best starting platform. There have been a number of approaches to this topic. Semsar-Kazerooni proposes a game theory approach to team cooperation, particularly team grouping [3]. Abraham has shown research into NPC team-mates
that assist the players [4]. In addition, psychological research examines the basis for human grouping and movement [5]. II. Approach In order to test these hypotheses, two teams of five s are placed on a map containing three s as shown in figure [1]. The goal of the game is to capture all three s. Alternatively, a team can win by eliminating all of the opposing team, as it is assumed the remaining team would then be able to freely capture all the s. Figure 1: A typical game setup In this game framework, various combinations of grouping and aggression are assigned to each team. Agents on a team can tend to group together or spread apart. They can tend towards making risky moves or attempt to remain in the relative safety of a. Finally, when given a choice, they can focus more on capturing s or eliminating enemies. III. The Agents Agents have multiple attributes that allow them to persist and interact within the game world. During the development of the s we wanted to make sure that the s were as individualistic and anonymous as possible to ensure that any results were affected by only changes within the itself. These attributes include: 1. Aggressiveness towards s versus s 2. How the is towards its target 3. Minimum and maximum distance to other teammates Any of these attributes when changed will affect the outcome of each simulation. The s two targets, s and s, make the choose a particular type of target with a much higher certainty. The minimum distance and the maximum distance restrict each s movement to a certain area around other s of its own team. This facilitates our version of grouping. In addition, the s have an armor rating, an attack rating, and a sighting distance. These values affect the s abilities to take damage, to deal damage, and to see targets. These values are not varied between s in these experiments. A. Agent Decisions The s group together by continually staying at a certain distance to and from the other s in its group. When s are set to stay in a group the maximum distance is set to 5 while the minimum distance is set to 3 which will restrict the s from moving to far from each other. However when the s are set to be non-grouping their maximum distance is set to be 25 with the minimum distance staying at 3. Having their maximum distance set to 25 allows the s to move freely throughout the map as individuals. The aggression of the s on each team is d off of an algorithm that determines where the will attempt to move. This algorithm is d on the notion of temperature. Agents in the system are attracted to higher temperatures. Each has a temperature that can fluctuate depending on the amount of activity at the. Each s temperature is d on its health. As each of the s is injured and has its health reduced its temperature goes up drawing other s towards it. The aggression algorithm is d off of the temperatures of both the and the s. The equation becomes the temperature divided by the distance plus the s corresponding aggression (temperature/distance)+aggression and the object that has the highest value becomes the s target. If an is deemed to be towards a certain object this value is increased by 1 if the target that the is evaluating is one of these objects. When an is, it is only towards s on the other team and s owned by the other team. If an is then it is not only towards s of the other team but also ones that are not owned by any team. In this context we deemed that an
Aggression constant of 1 as an appropriate number as it will not cause an to give up a very good opportunity to attack an object that it is not toward but is enough to cause it to move toward its aggressed target most often. If this algorithm were to be moved to a larger map with different temperatures this additional number would need to be changed to suit the environment. A typical aggression scenario is shown in figure [2]. draw s and players alike, together. In MMOs (Massively Multiplayer Online) and FPS games, s are used to increase the amount of people in a fight creating a much higher pace of action throughout the game. Without anything to draw players and s together the fighting becomes slow and there is a great reduction in the amount of fighting, thus reducing the appeal that the game has. Along with the s, the s also have attributes that affect the outcome of each simulation. These attributes include: 1. Radius 2. Temperature 3. Owner Figure 2 [In this figure the square team is while the triangle team is.] B. Agent Actions Movement the s have two main functions allowing movement, a rotate towards function and a move forward function. Each plans its movement, and then all s move at the same time, preventing either team from benefiting from additional knowledge. Attack When an comes within another s radius these s are allowed to attack one another. This ability to attack other s within the environment is the key ingredient to any action d game. When an attacks another, it only gets to attack one time toward one on each trip through the game loop. When each attacks, it removes the amount of attack it has from the enemy s armor amount. Take a As an moves around the map it will attempt to move toward s and s. When an comes within a s radius it will attempt to take the. When an takes a, it removes a certain amount of temperature from the adjacent. IV. Bases Bases represent another key attribute to many action d games. Bases create hot zones that tend to The radius of the defines how far away an has to be to be able to capture the. When a s temperature reaches the is considered taken and its owner becomes the team number of the that captured the. When all three s have the same owner the corresponding team wins. In order to draw s back to that the s temperature is increased by 1 each time the game loop progresses. The initial map set up included five s on each team placed just on the outside of the s on the right side of the map (Figure [1]). This style of map placement is traditional in both MMO style and also online FPS games during any sort of capture the style gameplay. The primary reason for placing the s like this is that it gives both teams easy access to a and then equal access to the third. This placement promotes a fair playing field to make sure that the map placement of the s did not affect the outcome of the simulations. To reinsure that this map placement did not affect the outcome of our simulations we rearrange the placement of the s giving each team a chance to start from either the top or bottom right s. This change produced near exactly the same results allowing for the conclusion that the map has no effect on the results. V. Data Each of the scenarios is divided into wins via killing off the other team and via capturing all of the s. All of the tested variables did make a significant impact on the outcome of the scenarios. In general when each team was given a more tendency towards a certain object, they tended to win according to the winning scenario of that object more often. Figure [3] depicts the wins d on Team 1 as and team as and both teams grouping.
Figure 3 [wins for teams that are grouping with ness towards s] When the teams are more towards s they tend to capture s more often than kill off the other team, the same can be said about when s are more towards other s. This tendency proves to us that the attributes in our when changed are having an effect on the outcome of our results. According to our results both and ness improves the teams chance of survival and victory. In general when s play against non s the s win more than times out of. 4 team = team = team 1 team team = Figure 4 [both teams grouping] team = team = team = In the starting hypothesis we decided that due to when s spread out they tend to attack as solo s and that each solo will become easily overwhelmed by the group of s. When the simulation was run, the data depicted in figure [6] on almost every account showed this exact result. The s that were grouping won more often than the s that were non-grouping. The only time that the non-grouping team won was when the non-grouping team was and the grouping team was. Thus pointing towards our hypothesis that it is good to stick together For each different set of attributes the simulation was run times in order to eliminate any outliers within the series. The attributes were tested with both teams set as grouping as in figure [4], both teams as non-grouping as in figure [5], and with one team as grouping and the other as non-grouping as in figure [6] (each of these figures depicts wins for each team pertaining to the scenario). In figure [4] the general outline of the data is even across the board besides when placing an team against a non- team. We believe that the reason for this is when a team groups up their ability to make it to all of the s is restricted because the s pull the whole team to one and then another and can t attack multiple s at a time thus making it an even playing field for wins. However when the teams group up they are able to kill other teams of s very efficiently, thus increasing their ability to win via killing the s on the other team. The difference in the amount of wins when run with two different kinds of ness can be attributed to the fact that our s have no survival instincts, meaning that if a team that is turns towards a it may decide to move towards the rather than returning fire. This would allow the s that are to kill the s with no resistance. team = team = Team 1 Team team = Figure 5 [both teams non-grouping] team = team = team =
In figure [5] the general outline of the data differs quite a lot from figure [4] there are 2 extremes in this data rather than just the one in figure [4] and, those are when you change both ness and ness. We believe that the change in ness is due to the same reason as when teams group up, being that the s tend to allow the s kill them, however when changing the ness in non-grouping s we see a large increase in wins. We believe that this is due to the fact that when non-grouping s are they divide and conquer and are able to take multiple s at a time rather than just one, this conclusion is d off of the fact that in our data the teams win by s almost 2 to 1 compared to winning with kills. In figure [6] we see the primary data, what happens when you change the way the teams group up and also their ness. We see a dramatic change in the amount of wins for the team that groups up. Our data shows that these extra wins are nearly always wins from killing the other team. We believe that this is due to the effect of swarming each that strays from its group is quickly overwhelmed by the team of s that is grouped up. VI. Future Work This project given the time limitations was not able to fulfill all of the desired outcomes. This project can take many new directions if given more work and time. 1. Advanced group decision making. At this point, only basic decisions are being made. It is possible that complex decisions such as having some team members assuming behavior while other team members remain behind to guard s could yield interesting results. 2. Complex survival instincts. In their current form, more s do not appear to have any benefits. If s could choose between ly pursuing a target versus taking cover or moving out of field of fire, survivability could be increased. 3. Enemy strategy discovery. Further work needs to be done to allow the team to determine which combination of parameters the enemy team is adopting in order to counter that strategy. VII. References 4 team = team = Team 1 Team 2 team = team = team = team = [1] Beij, A., and van der Sterren, W. (5). Killzone s AI: Dynamic Procedural Tactics Game Developer Conference, March 1, 5. [2] Semsar-Kazerooni, S. (June, 9). A game theory approach to multi- team cooperation. [3] Abraham, Aswin T. "AI for Dynamic Team- mate Adaptation in Games." (21): n. pag.ieee. Web. 4 Jan. 213. [4] Ruiz, Myriam A. "Team Agent Behavior Architecture in Robot Soccer." (n.d.): n. pag. IEEE. Web. 4 Jan. 213. [5] Derks, Belle. "Working for the Self or Working for the Group: How Self- Versus Group Affirmation Affects Collective Behavior in Low-Status Groups." Journal of Personality and Social Psychology 96.1 (9): 183-22. IEEE. Web. 4 Jan. 213. Figure 6 [Team 1 as grouping Team as non-grouping]