supplimentary materials

to start

the best way to start is to actually play the robot game. the task is to get the robots to guess the pattern (target) on the left by building a subset of the pattern in the middle (working area). you can click on the target pattern to sample a new pattern.

simply visit : https://evanthebouncy.github.io/projects/grids/

if you want to run a local version, open /interactive_mode/index.html in your favorite browser.

you can try some fun variants of the game where you get one of the robot to guess the pattern without revealing it to the other robot

contents

• README.md : this file

• version_space/ : the python code describing the DSL and generation of the version space caches of the meaning matrix

• interactive_mode/ : an interactive version where you can see the behaviours of both robots simultaniously

• user_study/ : code for study conducted on MTurk, complete with instructions, quiz, practices, trials, and survey

• user_study_results/ : the raw json results collected from the MTurk users used to make the plots

version_space/

• grid.py : the DSL, run it to sample a program and visualisation under version_space/drawings/ directory
• prag_cache.py : generate the version space cache for the meaning matrix, run it to generate a pickle L0VS.p. the generation is kind of jank, just leave it running for a few hours.
• prag1.py : the interactive mode in python, a prelim version before we build the javascript version in interactive_mode. requires L0VS.p to be precomputed
• ouput_json.py : translate L0VS.p to javascript equivalent jsons to be used in the javascript interactive systems. outputs all_shapes.js and l0vs.js.

interactive_mode/

there are many files here, but the important ones are:

• all_shapes.js : all the program executions as patterns (see version_space)
• l0vs.js : the version space cache (see version_space)
• L0.js : the literal listener code, requires all_shapes.js and l0vs.js.
• L1.js : the pragmatic code for both the speaker S1 and L1, requires L0.js.
• experiments.js and human_data.js : used to generate data for the plots of speaker-listener pairs (2nd figure in experiment section)
• index.* and assets/ : user interface code

user_study/

same as interactive_mode/ except with a user study flavor:

• index.js set up the random seeds for each user, including userID and the order of the stimuli, also contain the list of 10 representative patterns

user_study_results/

contains two json files of user study data, including time-stamped sequence of moves for each problem and the end-of-study survey