Introduction to datafsm"

Abstract

The datafsm package implements a method for automatically generating models of dynamic processes that both have strong predictive power and are interpretable in human terms. We use an efficient model representation and a genetic algorithm-based estimation process. This paper offers a brief overview of the software and demonstrates how to use it.

Introduction

This package implements our method for automatically generating models of dynamic decision-making that both have strong predictive power and are interpretable in human terms. We use an efficient model representation and a genetic algorithm-based estimation process to generate simple deterministic approximations that explain most of the structure of complex stochastic processes. The genetic algorithm is implemented with the GA package (Scrucca 2013). Our method, implemented in C++ and R, scales well to large data sets. We have applied the package to empirical data, and demonstrated the method's ability to recover known data-generating processes by simulating data with agent-based models and correctly deriving the underlying decision models for multiple agent models and degrees of stochasticity.

A user of our package can estimate models by providing their data in a common "panel data" format. The package is designed to estimate time series classification models that use a small number of binary predictor variables and move back and forth between the values of the outcome variable over time. Larger sets of predictor variables can be reduced to smaller sets with cross-validation. Although the predictor variables must be binary, a quantitative variable can be converted into binary by division of the observed values into high/low classes. Future releases of the package may include additional estimation methods to complement genetic algorithm optimization.

# Load and attach datafsm into your R session, making its functions available:
library(datafsm)

Please cite the package if you use it with the text generated by:

#library(printr)
citation("datafsm")

Fake Data Example

To quickly show it works, we can create fake data. Here, we generate 1000 repetitions of a ten-round game in which each player starts by making a random move, and in subsequent rounds, one player follows a "tit-for-tat" strategy while the other one follows a "noisy tit-for-tat" strategy that's equivalent to tit-for-tat, except that with a 10% probability the player will make a random move.

seed <- 1900
set.seed(seed)
cdata <- data.frame(outcome = NA,
                    period = rep(1:10, 2000),
                    my.decision1 = NA,
                    other.decision1 = NA)
#
# Prisoner's dilemma
#
pd_outcome <- function(player_1, player_2) {
  #
  # 1 = C
  # 2 = D
  #
  player_1  + 1
}

tit_for_tat <- function(last_round_self, last_round_opponent) {
    last_round_opponent
}

noisy_tit_for_tat <- function(last_round_self, last_round_opponent) {
  if (runif(1,0,1) <= 0.1) {
    sample(0:1,1)
  } else {
    last_round_opponent
  }
}

for (i in seq_along(cdata$period)) {
  if (cdata$period[i] == 1) {
    my.decision <- sample(0:1,1, prob = c(0.9,0.1))
    other.decision <- sample(0:1,1)
    cdata[i, "outcome"] <- pd_outcome(my.decision, other.decision)
  } else{
    my.last <- my.decision
    other.last <- other.decision
    my.decision <- tit_for_tat(my.last, other.last)
    other.decision <- noisy_tit_for_tat(other.last, my.last)
    cdata[i,c("outcome", "my.decision1", "other.decision1")] <- c(pd_outcome(my.decision, other.decision), my.last, other.last)
  }
}

The only required argument of the main function of the package, evolve_model, is a data.frame object, which must have 3-5 columns. The first two columns must be named "period" and "outcome" (period is the time period that the outcome action was taken). The remaining one to three columns are predictors, and may have arbitrary names. Each row of the data.frame is an observational unit, an action taken at a particular time and any relevant variables for that time. All of the (3-5 columns) should be named. The period and outcome columns should be integer vectors --- e.g. c(1,2,1) --- and the columns with the predictor variable data should be logical vectors --- e.g. c(TRUE, FALSE, FALSE) --- or vectors that can be coerced to logical with as.logical().

Here are the first eleven rows of this fake data:

knitr::kable(head(cdata, 11))

We can estimate a model with that data. evolve_model uses a genetic algorithm to estimate a finite-state machine (FSM) model, primarily for understanding and predicting decision-making. This is the main function of the datafsm package. It relies on the GA package for genetic algorithm optimization. We chose to use a GA because GAs perform well in rugged search spaces to solve integer optimization problems, are a natural complement to our binary representation of FSMs, and are easily parallelized.

evolve_model takes data on predictors and data on the outcome and automatically creates a fitness function that takes training data, an action vector that evolve_model generates, and a state matrix evolve_model generates as input and returns numeric vector whose length is the number of rows in the training data. evolve_model then computes a fitness score for that potential solution FSM by comparing it to the provided outcome in the real training data. This is repeated for every FSM in the population and then the probability of selection for the next generation is proportional to the fitness scores. If the argument cv is set to TRUE the function will call itself recursively while varying the number of states inside a cross-validation loop in order to estimate the optimal number of states, then it will set the number of states to that optimal number and estimate the model on the full training set.

If the argument parallel is set to TRUE, then these evaluations are distributed across the available processors of the computer using the doParallel functions; otherwise, the evaluations of fitness are conducted sequentially. Because this fitness function that evolve_model creates must loop through all the training data every time it is evaluated and we need to evaluate many possible solution FSMs, the fitness function is implemented in C++ to improve its performance.

res <- evolve_model(cdata, seed = seed)

evolve_model evolves the models on training data and then, if a test set is provided, uses the best solution to make predictions on test data. Finally, the function returns the GA object and the decoded version of the best string in the population. Formally, the function return an S4 object with slots for:

Use the summary and plot methods on the output of the evolve_model() function:

summary(res)
plot(res, action_label = ifelse(action_vec(res)==1, "C", "D"), 
     transition_label = c('cc','dc','cd','dd'))

The diagram shows that evolve_model recovered a tit-for-tat model in which the player in question ("me") mimics the last action of the opponent.

Use the estimation_details method on the output of the evolve_model() function:

suppressMessages(library(GA))
plot(estimation_details(res))

Documentation for evolve_model

Check out the documentation for the main function of the package to learn about all the options:

?evolve_model

Acknowledgements

This work is supported by U.S. National Science Foundation grants EAR-1416964 and EAR-1204685.

Session Info

sessionInfo()


Try the datafsm package in your browser

Any scripts or data that you put into this service are public.

datafsm documentation built on Nov. 17, 2017, 4:43 a.m.