Author: Sebastian Kranz, Ulm University
gtree is an R package that allows you to specify extensive form games using stages, similar as one specifies economic experiments with ztree or otree. The game is internally converted to a formal game tree and one can find its equilibria using a Gambit solver or an internal solver. The equilibria are presented in a format that facilitates comparison with experimental results. One can also conveniently study the effects of non-standard preferences characterized e.g. by inequality aversion or loss aversion.
Detailed documentation can be found on skranz.github.io/gtree.
To install gtree will all R dependencies simply run
install.packages("gtree",repos = c("https://skranz-repo.github.io/drat/",getOption("repos")))(My packages are hosted on a custom drat powered R repositorium. While CRAN ist great, I find it too time consuming to maintain all my packages there and devtools::install_github does not handle custom dependencies as easily.)
To effectively use gtree, you should also install Gambit from the Gambit project page. The easiest is to add the installation directory to your PATH, then gtree can directly use it.
Otherwise adapt the following call to globably specify the gambit directory:
options(gtree.gambit.dir="Your/gambit/path")
Here is an example of a game definition:
library(gtree)
game = new_game(
gameId = "Ultimatum_Dictator",
options = make_game_options(verbose=FALSE),
params = list(numPlayers=2,cake=10, autoAcceptProb=0.1),
stages = list(
stage("proposerStage",
player=1,
actions = list(
action("offer",~0:cake)
)
),
stage("autoAcceptStage",
nature = list(
natureMove("autoAccept",
set=c(TRUE,FALSE),
probs = ~ c(autoAcceptProb, 1-autoAcceptProb)
)
)
),
stage("responderStage",
player=2,
condition = ~ autoAccept == FALSE,
observe = c("offer","autoAccept"),
actions = list(
action("accept",c(FALSE,TRUE))
)
),
stage("payoffStage",
player=1:2,
compute=list(
payoff_1 ~ ifelse(accept | autoAccept, cake-offer,0),
payoff_2 ~ ifelse(accept | autoAccept, offer,0)
)
)
)
)This is a modified Ultimatum Game. Player 1 (the proposer) can decide how much of the parameter cake she offers to player 2 (the responder). With probability autoAcceptProb the responder has no choice and the offer is automatically accepted. Otherwise the responder can decide to accept or reject the offer. If he rejects, both players get a payoff of 0.
The game is defined by various stages, that can specify an active player and his actions, like offer, random moves of nature like autoAccept, or deterministically computed variables like payoff_1 or payoff_2 (the payoff variables have reserved names and must always be specified). In addition, one can explicitly specify which variables players observe in a stage and a condition specifying whether a stage is played or not. The corresponding fields can contain fixed values or formulas starting with ~ that reference to previously specified variables or parameters.
The following code assumes that players have inequality aversion preferences à la Fehr & Schmidt with an envy parameter of alpha=1 but no guilt (beta=0) and then solves for all pure strategy SPE of the game. The results in a table format:
game %>%
game_set_preferences(pref_ineqAv(alpha=1, beta=0)) %>%
game_solve() %>%
eq_tables()## $offer
## # A tibble: 1 x 2
## offer eq.inds
## <int> <chr>
## 1 4 1
##
## $accept
## # A tibble: 11 x 3
## offer accept eq.inds
## <int> <lgl> <chr>
## 1 0 FALSE 1
## 2 1 FALSE 1
## 3 2 FALSE 1
## 4 3 FALSE 1
## 5 4 TRUE 1
## 6 5 TRUE 1
## 7 6 TRUE 1
## 8 7 TRUE 1
## 9 8 TRUE 1
## 10 9 TRUE 1
## 11 10 TRUE 1
We see that the proposer offers 4 and every offer below 4 would be rejected. What happens for an auto accept probability of 90%?
game %>%
game_change_param(autoAcceptProb = 0.9) %>%
game_solve() %>%
eq_tables()## $offer
## # A tibble: 1 x 2
## offer eq.inds
## <int> <chr>
## 1 0 1
##
## $accept
## # A tibble: 11 x 3
## offer accept eq.inds
## <int> <lgl> <chr>
## 1 0 FALSE 1
## 2 1 FALSE 1
## 3 2 FALSE 1
## 4 3 FALSE 1
## 5 4 TRUE 1
## 6 5 TRUE 1
## 7 6 TRUE 1
## 8 7 TRUE 1
## 9 8 TRUE 1
## 10 9 TRUE 1
## 11 10 TRUE 1
While the responder's behavior does not change, the proposer will now only offer 0. Sure the offer will be rejected with 10% probability, but in 90% of cases the proposer gets the whole cake for herself.
Take a look at the different tutorials for more examples and features.
The companion package gtreeWebPlay allows to create shiny apps that allow a single player to play the game. The player plays against bots. A bot could e.g. follow an equilibrium strategy bot_eq or allow the current player to play against randomly drawn strategies of earlier users of the app bot_pop.
Currently there is no implementation for multiple human players to play against each other. To run controlled computerized experiments in which subjects directly play with, use e.g. ztree or otree instead.