func_beta: Function: Probability

In multiRL: Reinforcement Learning Tools for Multi-Armed Bandit

Function: Probability

Description

P_{t}(a) = \frac{ \exp(\beta \cdot (Q_t(a) - \max_{a' \in \mathcal{A}} Q_t(a'))) } { \sum_{a' \in \mathcal{A}} \exp( \beta \cdot (Q_t(a') - \max_{a'_{i} \in \mathcal{A}} Q_t(a'_{i})) ) }

P_{t}(a) = (1 - lapse \cdot N_{shown}) \cdot P_{t}(a) + lapse

Usage

func_beta(shown, qvalue, explor, rownum, params, hidden, system, ...)

Arguments

shown	Which options shown in this trial.
qvalue	The expected Q values of different behaviors produced by different systems when updated to this trial.
explor	Whether the agent made a random choice (exploration) in this trial.
rownum	The trial number
params	Parameters used by the model's internal functions, see params
hidden	All hidden variables within the MDP process belong here.
system	When the agent makes a decision, is a single system at work, or are multiple systems involved? see system
...	It currently contains the following information; additional information may be added in future package versions. idinfo: subid block trial exinfo: contains information whose column names are specified by the user. Frame RT NetWorth ... behave: includes the following: action: the behavior performed by the human in the given trial. latent: the object updated by the agent in the given trial. simulation: the actual behavior performed by the agent. position: the position of the stimulus on the screen. cue and rsp: Cues and responses within latent learning rules, see behrule state: The state stores the stimuli shown in the current trial—split into components by underscores—and the rewards associated with them.

Value

A NumericVector containing the probability of choosing each option.

A List

• output [NumericVector]

  A numeric vector representing the probability of selecting each option.

  The inverse temperature parameter beta in the softmax function primarily controls these probabilities. Larger values of beta make choices more sensitive to differences in Q-values, while smaller values make choices closer to random.

  In addition, the lapse parameter prevents the probability of any option from reaching zero, thereby avoiding logP = -Inf.

• hidden [CharacterVector]

  User-defined internal variables generated by this function. These represent intermediate (latent) states produced during computation, which can be read or modified by other functions in the MDP process.

Body

func_beta <- function(
    shown,
    qvalue, 
    explor,
    system,
    rownum,
    params,
    hidden,
    ...
){
  
  list2env(list(...), envir = environment())
  
  # If you need extra information(...)
  # Column names may be lost(C++), indexes are recommended
  # e.g.
  # Trial  <- idinfo[3]
  # Frame  <- exinfo[1]
  # Action <- behave[1]
  
  beta     <- params[["beta"]]
  lapse    <- params[["lapse"]]
  weight   <- params[["weight"]]
  capacity <- params[["capacity"]]
  
  index     <- which(!is.na(qvalue[[1]]))
  n_shown   <- length(index)
  n_system  <- length(qvalue)
  n_options <- length(qvalue[[1]])
  
  # Assign weights to different systems
  if (length(weight) == 1L) {weight <- c(weight, 1 - weight)}
  weight <- weight / sum(weight)
  if (n_system == 1) {weight <- weight[1]}
  
  # Compute the probabilities estimated by different systems
  prob_mat <- matrix(0, nrow = n_options, ncol = n_system)
  
  if (explor == 1) {
    prob_mat[index, ] <- 1 / n_shown
    prob_mat[prob_mat == 0] <- NA
  } else {
    for (s in seq_len(n_system)) {
      sub_qvalue <- qvalue[[s]]
      exp_stable <- exp(beta * (sub_qvalue - max(sub_qvalue, na.rm = TRUE)))
      prob_mat[, s] <- exp_stable / sum(exp_stable, na.rm = TRUE)
    }
  }
  
  # Weighted average
  prob <- as.vector(prob_mat %*% weight)
  
  # lapse
  prob <- (1 - lapse * n_shown) * prob + lapse
   
  return(list(output = prob, hidden = hidden)) 
}

multiRL documentation built on June 9, 2026, 5:09 p.m.