Science

In addition to being an original improv comedy show, Improbotics is also the subject of scientific research on the use of artificial intelligence in theatre and on the Turing test*, and it has been presented at international scientific conferences.

Piotr and Kory, who are both improvisers and researchers in AI, believe in an interdisciplinary approach to the arts and sciences. Our vision is to explore new forms of human theatre through the help of state-of-the-art research in artificial intelligence and robotics, and to highlight what is uniquely human. For us, the skill of theatrical improvisation is one of the highest forms of human intelligence, and by confronting humans to the incongruity of machines on the stage, we can create new opportunities for comedy.

We are planning on continuing writing about the science behind AI-based improv comedy stage partners. The experimental protocol involves a Turing test* that takes place in a theatre, with both the audience members and the performers left to guess who is a human and who is a machine. In order to test scientific hypotheses about the perception of humans vs. machines and to conduct the Turing test, we typically collect anonymized feedback from audience members as well as feedback from volunteer performers who signed consent forms. Participation in the data collection is always be optional and never mandatory. The experimental protocol has been validated by the institutional ethics board of the University of Alberta. Improbotics performers and artistic collaborators are invited to contribute as authors to the paper if they wish, and all artists are always acknowledged.

* To pass the Turing test, a machine (such as an artificial intelligence-based chatbot) needs to display the appearance of human-level intelligence by fooling a human judge into thinking that it is human.

Technology

At the heart of our improv show is the piece of technology that Piotr Mirowski and Kory Mathewson have developed over the years, as part of HumanMachine, and consisting in an artificial intelligence-based chatbot, called A.L.Ex.

A.L.Ex, which stands for Artificial Language Experiment (and is also an obscure reference to the clever parrot), is a computer system that can do speech recognition, improvised dialogue and voice synthesis. It is powered by recurrent neural network software, trained on movie dialogue from a hundred thousand movies, and more recently, 4GB of web pages, and it is programmed by Piotr and Kory using open-source components (custom software build on top of torchrnn trained on movie dialogue from OpenSubtitles, GloVe word embeddings, Vowpal Wabbit, OpenAI‘s GPT-2, Google Research Universal Sentence Encoder). Despite being state-of-the-art, A.L.Ex is still a weak form of artificial intelligence.

At its core, A.L.Ex is a language model that captures the complex statistics of  words. It is trained to produce the most obvious words that follow any context. It illustrates the fundamental principle of improv: always do the most obvious thing.

A.L.Ex runs on a laptop computer. In the case of HumanMachine’s show, A.L.Ex’s stage presence consists in a portable microphone, amplified sound and video projection, and in a robot produced by EZ-Robot, for which we developed custom control software. In the case of Improbotics shows, A.L.Ex controls human improvisers by sending them speech synthesized lines of dialogue via headsets and FM radios.