Proc. the 34th Annual conference of International Simulation And Gaming Association (ISAGA2003), Chiba, Japan, pp. 893-900, August 25-29, 2003.
Identification of Player Types in Massively Multiplayer Online Games Ruck Thawonmas, Ji-Young Ho, and Yoshitaka Matsumoto Intelligent Computer Entertainment Laboratory Department of Computer Science, Ritsumeikan University Kusatsu, Shiga 525-8577, Japan
[email protected]
Introduction In this paper, we discuss an approach for identification of player types in massively multiplayer online games (MMOGs). MMOGs provide fast growing online communities (Jarett et al. 2003). Managing a large-scale community implies many challenges, such as identification of player types, social structures, and virtual economic mechanisms, etc. In this paper, we address the challenge on identification of player types. As a first step toward use of real MMOG data, we demonstrate our approach using a PC1 Cluster based MMOG simulator. The work presented in this paper is divided into two phases, namely, modeling and identification. In the modeling phase, many types of player agents with different characteristics are modeled using the above MMOG simulator. By player agents, we mean agents that imitate player characters in real MMOGs. The player agents reside in and migrate among multiple game worlds, each world running on a PC node. A game world also accommodates monsters, representing no player characters in real MMOGs, that could kill (or be killed by) player agents. In the identification phase, the task is to correctly identify the type of a given player agent merely from its log. The information on the agent models is not used here, assuming that such information is hardly available beforehand in real MMOGs. To perform this task, two technical issues are discussed. The first one is feature selection, namely, selection of input features from log data. The other one is classifier selection, namely, selection of a classifier for identifying a given player agent to a particular type based on the selected input features. 1
In this paper, PC stands for personal computer, and should not be confused with player character.
2
Ruck Thawonmas, Ji-Young Ho, and Yoshitaka Matsumoto
Fig. 1. Zereal architecture.
The organization of the rest of the paper is as follows: MMOG Simulator and Agent Modeling section gives a description on the PC Cluster based MMOG simulator that we use, followed by a description on agent modeling. Player Identification section discusses feature selection and classifier selection. Experiments section describes the experiments and gives the experimental results. Finally, Conclusions section concludes the paper by summarizing the main results and suggesting future work.
MMOG Simulator and Agent Modeling The PC Cluster based MMOG simulator that we use is Zereal (Tveit et al. 2003). Zereal is one of multi agent simulation systems (MASs) (Epstein and Axtell 1996). It can simulate multiple game worlds simultaneously running each game world on a different PC node. Fig. 1 shows the architecture of Zereal. Zereal is composed of one master node and multiple world nodes. The master node collects the current status (world model) of each game word and forwards this information to a client computer for visualization or data analysis. A world node simulates all game objects such as player agents and monster agents. Other game objects include food items and potion items for recovering stamina, and key items for opening a door in order to leave the current game world. In the version of Zereal that we licensed from the Zereal developing team, three types of player agents, namely, Killer, Markov Killer, and Plan Agent, are provided. Each type has 7 common actions, namely, Walk, Attack, Pick-
Identification of Player Types in Massively Multiplayer Online Games
3
Table 1. Relative frequencies (columnwise) of player agent actions. PC Types Walk Attack PickFood PickPotion PickKey LeaveWorld Killer L H M M L L Markov Killer M M H H M M Plan Agent H L L L H H
Fig. 2. Screen shot of ZerealViewer.
Food, PickKey, LeaveWorld, and EnterWorld (not used in our study), but is designed to have a different behavior described as follows: • Killer puts the highest priority on attacking monsters. • Markov Killer selects the next action according to the corresponding state-transitional probability. • Plan Agent focuses on finding of a key to leave the current game world. Table 1 shows the relative frequencies of player agent actions compared to the others. In the table, each element is compared columnwise, and H, M, L stand for high, medium, and low, respectively. To observe activities in the artificial societies, visualization tools are crucial for MASs. We have developed such a tool called ZerealViewer. Though not yet fully functioned, a screen shot of the ZerealViewer when one game world is simulated is shown in Fig. 2. Fig. 3 shows a typical game log sent to the client from the master node for data analysis. The first and the second columns in the log indicate the simulation time steps and the real clock time, respectively. The third one shows the agent identifier numbers with the most upper digit(s) indexing the current world node. The fourth column represents agent actions, and the fifth
4
Ruck Thawonmas, Ji-Young Ho, and Yoshitaka Matsumoto
Fig. 3. Typical game log.
and sixth columns the coordinates in the game world before and after such actions, respectively. The last column gives information on the types of agents.
Player Identification The task here is that of identifying the type of a given player agent merely from its log. In our case, though type information is already available in the log, this information is not used. Feature Selection From Fig. 3, a sequence of actions can be easily achieved for each player agent. Exploiting this information, we hence propose the following procedure for deriving the input features. • Step I For each player agent, sum up the total number of each action that the player agent performed. • Step II For each player agent, divide the result of each action in Step I by the total number of all actions that the player agent performed. (Input Feature Set A)
Identification of Player Types in Massively Multiplayer Online Games
5
Table 2. Typical results of Step I. PC Types Plan Agent 1 Killer 1 Killer 2 Markov Killer 1 Markov Killer 2 Plan Agent 2
Walk Attack PickFood PickPotion PickKey LeaveWorld 113 0 0 0 8 1 67 92 2 0 0 0 93 104 0 1 2 0 107 1 6 2 0 0 177 11 4 7 1 0 119 0 1 0 4 0
Table 3. Typical results of Step II. PC Types Plan Agent 1 Killer 1 Killer 2 Markov Killer 1 Markov Killer 2 Plan Agent 2
Walk 0.9262 0.4161 0.4650 0.9224 0.8850 0.9597
Attack PickFood PickPotion PickKey LeaveWorld 0 0 0 0.0656 0.0082 0.5714 0.0124 0 0 0 0.5200 0 0.0050 0.0100 0 0.0086 0.0517 0.0172 0 0 0.0550 0.0200 0.0350 0.0050 0 0 0.0081 0 0.0323 0
Table 4. Typical results of Step III. PC Types Plan Agent 1 Killer 1 Killer 2 Markov Killer 1 Markov Killer 2 Plan Agent 2
Walk 0.9651 0.4336 0.4845 0.9612 0.9222 1.0000
Attack PickFood PickPotion PickKey LeaveWorld 0 0 0 1.0000 1.0000 1.0000 0.2402 0 0 0 0.9100 0 0.1429 0.1525 0 0.0151 1.0000 0.4926 0 0 0.0963 0.3867 1.0000 0.0762 0 0 0.1559 0 0.4919 0
• Step III For each player agent, divide the result of each action in Step II by that of the player agent who most frequently performed the action. (Input Feature Set B) In the experiments, it will be shown that Input Feature Set B outperforms Input Feature Set A in more realistic environment. Tables 2, 3, and 4 respectively show examples of typical results of Steps I, II, and III for a log taken during 200 simulation time steps. Note that in Table 2 the rowwise sum for each player agent need not be 200 since a player agent may not perform any action during some simulation time steps.
6
Ruck Thawonmas, Ji-Young Ho, and Yoshitaka Matsumoto
Fig. 4. Concept of MBR on 2D data.
Classifier Selection In our work, we adopt memory-based reasoning (MBR) (Berry and Linoff 1997) as the classifier. Our main reason is that setting of initial parameters in MBR requires less effort compared to other classifiers. The algorithm of MBR is also simple, namely, for a given unknown data, find k nearest neighbors in input feature space among the training data, and then take the majority vote to decide the type of the unknown data. In MBR, thereby, we need to decide only k and the distance measure. Fig. 4 shows the concept of MBR on 2-dimensional data, when k = 3 and Euclidian distance is used. In this figure, we assume that there are 2 types of data, depicted respectively by • and ◦. The identifying result for unknown data × is that it is of type •.
Experiments Any classifier should be able to correctly identify unknown data not seen in training data. This ability is called generalization ability. To approximate the generalization ability, we use leave-one-out method (Weiss and Kulikowski 1991). In leave-one-out method, supposing that the total number of available data is M , first, data number 1 is used for testing and the other data are used for training the classifier of interest. Next, data number 2 is used for testing and the other data are used for training the classifier. The process is iterated in total M times. In the end, the averaged recognition rate for test data is computed, and is used to indicate the generalization ability of the classifier. For experiments, log data were obtained by running 5 game worlds simultaneously, each game world having 5 player agents of each type, 5 monsters, and 5 items for each of the other game objects. Migration of player agents
Identification of Player Types in Massively Multiplayer Online Games
7
Table 5. Generalization ability of MBR with k = 1. Input Feature Set N0 N1 N2 A 97% 60% 51% B 97% 85% 69%
Table 6. Generalization ability of MBR with k = 5. Input Feature Set N0 N1 N2 A 88% 63% 54% B 96% 88% 74%
Table 7. Generalization ability of MBR with k = 9. Input Feature Set N0 N1 N2 A 89% 64% 54% B 93% 89% 77%
among game worlds was prohibited. For the log data, we conducted the feature selection procedure discussed in Feature Selection subsection, and obtained in total 75 data for Input Feature Set A and 75 data for Input Feature Set B. In real MMOGs, players sometimes might select different actions even when they face the same situations. To make our experiments more realistic, we added Gaussian noises to the above data. We did this with 3 levels of noises, namely, • N0 Gaussian noises with mean = 0 and variance = 0.001. • N1 Gaussian noises with mean = 0 and variance = 0.1. • N2 Gaussian noises with mean = 0 and variance = 0.2. Tables 5, 6, and 7 respectively show the generalization ability of MBR using Input Feature Set A and Input Feature Set B with k = 1, 5, 9, respectively. As can be seen from these tables, Input Feature Set B outperforms Input Feature Set A, especially when the levels of noises are high. This is because Step III in Feature Selection subsection increases the value of each element in each column. For example, notice the changes in column PickFood in Table 3 and Table 4. It is clear that elements with high values are less affected by noises than those with low values.
8
Ruck Thawonmas, Ji-Young Ho, and Yoshitaka Matsumoto
Conclusions In this paper we have presented an approach for identification of player types in MMOGs. We proposed a procedure for selection of features that are robust against noises (Input Feature Set B). With this set of features, MBR, adopted as the classifier, could successfully identify the type of unknown data in a low noise case. For cases with higher noise levels, the performance of MBR using Input Feature Set B drops more gracefully than that of MBR using the non-robust input feature set (Input Feature Set A). In the future work, we are going to conduct experiments using agents with more complicated behaviors and try on some other classifiers.
Acknowledgements This work has been supported in part by the Ritsumeikan University’s Kyoto Art and Entertainment Innovation Research, a project of the 21st Century Center of Excellence Program funded by the Japan Society for Promotion of Science. The second author has been supported by a scholarship from the Ministry of Education, Culture, Sports, Sciences and Technology, Japan.
References 1. Berry M, Linoff G (1997) Data Mining Techniques-For Marketing, Sales, and Customer Support. John Wiley & Sons, Inc., New York 2. Epstein J, Axtell R (1996) Growing Artificial Societies: Social Science from the Bottom Up. MIT Press, Brookings, MA 3. Jarett A, Estanislao J, Dunin E, MacLean J, Robbins B, Rohrl D, Welch J, Valadares J (2003) IGDA Online Games White Paper. 2nd Edition 4. Tveit A, Rein O, Iversen J, Matskin M (2003) Zereal: A Mobile Agent-based Simulator of Massively Multiplayer Games. under submission 5. Weiss S, Kulikowski C (1991) Computer Systems That Learn. Morgan Kaufmann Publishers, San Mateo, CA
