Mouselab-MDP

Mouselab-MDP is a JsPsych plugin that renders Markov decision processes (MDP) for participants to navigate through. Optionally, information about the environment can be presented only when the participant asks for it, thus creating a record of the information considered. Try out the live demo mouselab-demo here.

Usage

To view the experiment, you must use an http server. You can start one by running python -m http.server in the experiment/ directory. Then visit http://localhost:8000/ in your browser.

A basic description of how to use the plugin can be found in our RLDM paper . There is also a limited amount of inline documentation in the source coffeescript. Pay special attention to the constructor of MouselabMDP, beginning at line 90.

Note that we have included the javascript which compiles from coffeescript. You only need to link to the javascript in your own code (I just include the coffeescript for reference).

Example

Documentation is sparse, thus it is probably best to learn by example. The code for the demo can be found in experiment/. Here is an excerpt from the main experiment file, the definition of a simple Mouselab-MDP trial.

trial =
  type: 'mouselab-mdp'  # use the jspsych plugin
  # ---------- MANDATORY PARAMETERS ---------- #
  graph:  # defines transition and reward functions
    A:  # for each state, an object mapping actions to [reward, nextState]
      right: [5, 'B']  # states may be strings or numbers (be consistent!)
      left: [-5, 'C']
    B: {}
    C: {}
  layout:  # defines position of states
    A: [1, 1]
    B: [2, 1]
    C: [0, 1]
  initial: 'A'  # initial state of player
  # ---------- OPTIONAL PARAMETERS ---------- #
  stateLabels:  # object mapping from state names to displayed labels
    A: 'A'
    B: 'B'
    C: 'C'
  stateDisplay: 'always'  # one of 'never', 'hover', 'click', 'always'
  stateClickCost: 0  # subtracted from score every time a state is clicked
  edgeLabels: 'reward'  # object mapping from edge names (s0 + '__' + s1) to labels
  edgeDisplay: 'hover'  # one of 'never', 'hover', 'click', 'always'
  edgeClickCost: 0  # subtracted from score every time an edge is clicked
  stimId:1994  # for your own data-keeping
  playerImage: 'static/images/plane.png'
  playerImageScale: 0.5
  size: 200  # determines the size of states, text, etc...
  leftMessage: 'Left Message'
  centerMessage: 'Center Message'