Human Strategic Steering Improves Performance of Interactive Optimization

Instructions

Before running the demo, you can learn here how to use the system we used in our study. Publication Preprint

This demo has been slightly modified from the original to run entirely in a modern web browser using Pyodide to use the required Python code.

Note however that Pyodide does not support Safari (on any platform) and has been reported to not work on any iOS (link). In general we recommend to use run our demo on a laptop using the latest stable version of Google Chrome or Firefox, which are known to work.

Task

The task you need to accomplish is to guide the AI, using your feedback, to find the maximum of a function (i.e. the AI queries the highest peak of the function) in the shown interval, with the least number of iterations.

Setting

The AI you will collaborate with, is smart but blind. This means that the only thing the AI can see is your feedback, while only you will be able to see the function (the solid line ──). Therefore it needs your help to find the maximum. You will also be able to see what the AI thinks: both its estimate of the function (dashed line ╍╍) and its uncertainty (blue shadow ) are visualised.

Interaction

During each iteration the AI will try to guess on which point, along the x-axis, the maximum lies, and it will query it to you, asking for your feedback about the y-value of the function in that queried point. You can move your cursor up and down over the plot, and then answer by clicking at the height that you consider the most appropriate. The AI uses your feedback to figure out which point to query next, so the value you choose is extremely important. You are free to answer the exact height of that point, or to tell the AI any other value, if you think it will help the AI more. If the AI queries the maximum, try to keep it querying on that same point.

Score

Your score is not the height of your feedback, but the height of the highest point the AI queried (indicated by a red dot ●). It ranges between 0 and 100 (0 at the minimum of the function, 100 at its maximum). The initial query is random, so your initial score is random as well.

Configurations

Upon beginning you can select different configurations:

• default default settings, for demo purposes. Features 5 practice sessions and 5 normal sessions, each of 5 iterations;

• study_training used for training before the actual study. Features 5 practice sessions each of 10 iterations;

• study_10 used in the study. Features 5 practice sessions and 15 normal sessions, each of 10 iterations;

• study_5 used in the study. Features 5 practice sessions and 20 normal sessions, each of 5 iterations.