aalto-ui/MayAI

May AI? Design Ideation with Cooperative Contextual Bandits

In the May AI-Project we developed a Cooperative Contextual Bandits (CCB) for an interactive design ideation tool that 1) suggests inspirational and situationally relevant materials (“may AI?”); 2) explores and exploits inspirational materials with the designer.

In order to do so, we adapted an online learning cooperative contextual bandit algorithm presented by Tekin et al. to address our objectives to 1) identify and propose material that is novel and relevant for a designer yet also 2) adapt to the designer’s changing strategy of diversification and intensification.

Overview of the Algorithm:

We first slice the potential mood board space defined as: color value (C), saturation (S), color lightness (L), image orientation (O), and color distance (D) into partitions (see bandit.py). These are then handled by strategy agents (A), where each is responsible for recommendations by its suggestion agents (a).

In every discrete trial:

a mood board is transformed into a five-dimensional vector in the context space and is assigned to the strategy agent of the corresponding partition;
the agent queries its own suggestion agents for similar suggestions (exploitation) and nearby strategy agents for alternative moods (exploration);
each suggestion agent within the current strategy, and each nearby strategy agent, provides probabilities for making a good suggestion (Fig. 1: b);
the agent with the highest probability is selected with respect to an exploration/exploitation criterion, c;
if a suggestion agent is selected (Fig. 1: c), it describes the next image suggestion feature vector; otherwise (Fig. 1: d), the corresponding strategy agent queries its own suggestion agents to identify this vector;
this vector, in combination with the association list, is used to query a suitable image in the local database; if not successful, it will be translated into humanreadable features to query images online in real time;
the user accepts or rejects the suggested image; and
that feedback is used to update the probability distributions of the corresponding suggestion agents and, in case of referral, of the neighboring strategy agent.

The code for step 1 - 5 will be provided in the following.

Getting started:

Dependencies

pip install functools
pip install scipy
pip install random2
pip install numpy

Python version

2.7

Usage:

Navigate to the folder. Run:

# Install required packages
python main.py

It will ask for an initial image in the form of a vector: (Hue:0-360, Saturation:0-1, Lightness:0-1, Orientation:0,1, Color Distance between the most dominant colors: 0-180)
An example vector would be: (123,0.6,0.3,0,178)
The CCB will then suggest another image vector, which can be accepted (y) or rejected (n) or can be stopped (quit).

License

This project is licensed under the MIT License, see the LICENSE.txt file for details.

aalto-ui / MayAI