Utils

Module containing functions that are used across Fitr modules

References

[R98b6f8855756-Akam2015]

Akam, T. et al. (2015) Simple Plans or Sophisticated Habits? State, Transition and Learning Interactions in the Two-Step Task. PLoS Comput. Biol. 11, 1–25

Module Documentation

fitr.utils.action(x)

Selects an action based on state-action values

Parameters:

x : ndarray

Array of action values (scaled by inverse softmax temperature).

Returns:

int

The index corresponding to the selected action

Notes

This function computes the softmax probability for each action in the input array, and subsequently samples from a multinomial distribution parameterized by the results of the softmax computation. Finally, it returns the index where the value is equal to 1 (i.e. which action was selected).

fitr.utils.logsumexp(x)

Numerically stable logsumexp.

Parameters:

x : ndarray(shape=(nactions))

Returns:

float

Notes

The numerically stable log-sum-exp is computed as follows:

\[\max X + \log \sum_X e^{X - \max X}\]

fitr.utils.softmax(x)

Computes numerically stable softmax

Parameters:

x : ndarray(shape=(nactions))

Returns:

ndarray(shape=(nactions))

Softmax probabilities for each action

fitr.utils.trans_UC(values_U, rng)

Transforms parameters from unconstrained to constrained space

Parameters:

values_U : ndarray

Parameter values

rng : {‘unit’, ‘pos’, ‘half’, ‘all_unc’}

The constrained range of the parameter

Returns:

ndarray(shape=(nparams))

Notes

This code was taken from that published along with [Akam2015].