The Werewolf game is a conversation-based party games. Each player in the game belongs to werewolves or villagers. Since secret conversations called \Whispers" are allowed for werewolves only, effective use of whispers must be a key issue for werewolves to proceed advantageously to win the game. In this work-in-progress paper, for a preliminary assessment of the whispers' effect, we extract relational association rules having behaviors in whisper from the log data of Werewolf BBS.
In recent years, there has been a growing interest in the research of Arti cial
Intelligence. Current technologies in AI reach a level high enough to beat
human in complete information games such as Shogi and Go. As a next step for
realizing general arti cial intelligence, incomplete information games are
receiving increased attention. As one of representative incomplete information games,
the Werewolf game is widely recognized as promising research testbed for
intelligent agents in Japan, and a project for making arti cial intelligence based
Werewolf (AIWolf)1 is established recently. Intensive researches are conducted
from various aspects for realizing AIWolf, e.g. [
The Werewolf game is a conversation-based party game which models a conict between werewolves who are minorities having rich information and villagers who are majorities having less information. There exist two types of conversations in the game. One is an open conversation, and the other is a closed or secret conversation. While all alive players in the game can join and browse the open conversations, secret conversations are allowed for werewolves only. Thus, effective use of secret conversations must be a key issue for werewolves to proceed advantageously to win the game.
In this work-in-progress paper, we focus on information differences between
werewolves and villagers, and try to capture a characteristic relationship between
contents in the secret conversations and actual utterances in the open
conversations. For this purpose, we extract relational association rules[
The Werewolf game is a multiplayer communication party game. Each player belongs to werewolves side or villagers side. A werewolf player knows who belong to the same side, but villagers have no information on other players' side. Some villager has a special ability. Seers can know that the designated player is a werewolf or not. Mediums can know that an executed player was a werewolf. Hunter can guard a designated player from the attack by werewolves. The game has two phases, daytime phase and nighttime phase, to be iterated. In daytime phases, all players join the open conversation and give vote for deciding an executed player. In the conversation, villagers try to nd out werewolves and werewolves try to deceive villagers. In nighttime phase, werewolves select a dangerous villager and attack him/her. Executed or attacked players are exiled from the game. Villagers win the game if all werewolves are executed, while werewolves win if the number of villagers is no more than that of werewolves.
The Werewolf BBS2 is an online BBS website for playing text-based Werewolf games. The rules in the BBS are almost the same as those in the original Werewolf games with a few exceptions. The BBS has four types of log data storing players' utterances. A \white log" stores all utterances during the open conversations. All players can browse a white log. A \red log" keeps the utterances called \whispers" in a secret conversation among werewolves. We employ these two kinds of log data for the analysis. 2.2
Each utterance is written in natural language. To extract essential meanings of
utterances and convert them machine manageable, a communication protocol
for the Werewolf game is proposed in [
Player2 a question on the Dayth day in a game Game. 2 http://www.wolfg.x0.com/ w request divine( Game:Day, Player, Player2 ) : A player Player requests seers to divine the team which a player Player2 belongs to on the Dayth day in a game Game.
A complete list of red predicates is shown in Table 1. We explain a couple of red predicates below. r want eat( Game:Day, Player, Players2 ) : A werewolf Player wants to attack a player Player2 on the Dayth day in a game Game. r estimate( Game:Day, Players, Player2, Role) : A werewolf Player estimates that a player Player2 has a role of Role on the Dayth day in a game Game. r decieve( Game:Day, Player, Player2, Role) : A werewolf Player offers a werewolf Player2 to behave as Role to deceive villagers on the Dayth day in a game Game.
Three arguments, Game, Day and Player are in common in all predicates for handling a chain of utterances. In addition, to relate the past utterances to the current one, a rule
Pred( Game:Day, N, Player, Args ) :
prev days(N), PDay is Day-N, Pred( Game:PDay, Player, Args ). is employed for each predicate, in which a predicate prev days(N) returns a nonnegative integer N. This rule states that a player Player took an action Pred N days ago from Dayth day in a game Game. 3 3.1
We select six games from the Werewolf BBS. All of them have twelve villagers and three werewolves including at least deceiving one. Werewolves won three of six games, and lost the rest three.
All white and red logs are manually converted into the predicates introduced in the previous section. The average numbers of facts on red predicates over three games the werewolves won and lost respectively are summarized in Table 1. From the table, we can con rm that the main topics in secret conversations are question, answer, advice, estimate, and want eat. Furthermore, each number in the games werewolves won is more than that in the games werewolves lost, even if we consider the number of werewolves executed. In other words, intensive communications are observed in the game werewolves won. 3.2
In this work-in-progress paper, relational association rules to be extracted are restricted to have at least one red predicate in their body. Furthermore, they have to contain one of three head predicates below:
Note that, since werewolves know who werewolves are, designating villager as a werewolf in the predicate wolves estimate wolf indicate that a werewolf tries to deceive other villagers. Werewolves may also designate a werewolf to avoid a suspicion. We extract facts on the above three head predicates from log data. As a result, 27, 39 and 53 facts are obtained for attacked, executed and wolves estimate wolf, respectively.
Three interestingness measures are used for evaluating relational association rules. The rst one is support count which is de ned as a number of distinct instantiations of head variables by which we can derive both of head and body. The second one is con dence value or conditional probability. It is de ned as a probability that an instantiation of head variables satisfying the body can derive the head. To assess the rough effects of the red predicates in the whole, we employ the third measure D = P (Head j Body) P (Head) where P (Head j Body) is the con dence value and P (Head) is a priori probability that the head holds. The value of P (Head) is estimated by using all possible instantiations of head predicate considering alive players and their roles. The positive value of this measure indicates that the body predicates have positive effects to the head, while negative one shows the negative effect of the body. 3.3
An inductive logic programming engine Aleph3 is employed to extract all association rules regardless of that they contain red predicates or not. We give Aleph system a certain parameter setting for association rule search and execute it with the induce max command. Relational association rules satisfying our conditions are extracted from the results of Aleph system in a post-processing. As a result, 4702, 3048 and 3094 rules are obtained having the predicate attacked, executed and wolves estimate wolf, respectively.
A couple of derived association rules having high con dence value are shown below. 1. Werewolves attack a player C whom werewolves want to attack if C asked a question for a player E estimating a relationship between two players. attacked( Game:Day, C ) :r want eat( Game:Day, 0, D, C ), w question( Game:Day, 1, C, E ), w line( Game:Day, 1, E, F, G ). 2. A player C is executed if C is given a vote by a player D whom the werewolves estimated as hunter. executed( Game:Day, C ) :w vote( Game:Day, 0, D, C ), r estimate( Game:Day, 2, F, D, hunter ). 3. A werewolf state that a player C is a wolf if a player E whom werewolf F wants to attack agree with C. w estimate wolf( Game:Day, C ) :w agree( Game:Day, 2, E, C), r want eat( Game:Day, 2, F, E ). predicate
As one of future works, we plan to extract condensed representations of
relational association rules[