Adapting User Interfaces with Model-based Reinforcement Learning (ACM CHI 2021)

By Kashyap Todi, Gilles Bailly, Luis A. Leiva, Antti Oulasvirta

https://userinterfaces.aalto.fi/adaptive

Copyright (c) 2021 Aalto University. All rights reserved.

This code repository is for the adaptive menus application described in the CHI 2021 paper: https://userinterfaces.aalto.fi/adaptive/resources/chi2021-todi-adaptive.pdf

Requirements

1. Ray (https://ray.io) for parallelisation.
2. TensorFlow 2 (https://www.tensorflow.org/install/pip) for using the neural networks.

Key Components for the Menu Adaptation Application

• plan.py is the starting point for code execution. To generate results, execute:

• python3 plan.py The command will run the MCTS planner for the 5-item case (menu_5items.txt) without the value network. To use the value network, add the -nn option.

• python3 plan.py -h See the full list of options available for running plan.py

• utility.py contains useful functions for loading data, initialisation, etc.

• Input files are stored within Input folder. For each case, there's an input menu, association list, and user history.

• state.py defines the menu and user state. The root state is initialised using the input menu, associations, and user history

• adaptation.py provides a general format for defining adaptations. It uses the syntax (i,j,type,expose) where i and j are two positions in the menu, type specifies the type of adaptation (e.g. swap, move, group move), and expose is a boolean that specifies whether the adapted menu is exposed to the user or not.

• mcts.py contains the code for Monte Carlo tree search.

• useroracle.py defines the user models for running simulations. These models are used towards predicting task completion time given a menu design, and for computing the reward after making an adaptation.

Read the Value Network documentation.

Read the Policy Network documentation.