Vivarium Squirrel Logo: Genetic Algorithms and the Vivarium Media Lab Project

In 1985, I had the opportunity to attend an intriguing class at the MIT Media Lab. It was a collaborative journey with my friend Noyuri Mima, who was then a graduate student at the Harvard Graduate School of Education, while I was serving as a research staff member at the MIT AI and LCS Laboratory under Hal Abelson and Gerry Sussman

The course, named 'The Vivarium', was officially led by Alan Kay and Stewart Brand. Though Alan Kay's appearances were rare, Stewart Brand and Brian Harvey, a consistent lecturer, predominantly conducted the sessions.

Brian Harvey, notable for his work on Logo during his undergraduate years at MIT and later celebrated for developing high school computer science curricula based on Scheme at UC Berkeley, introduced us to an ambitious concept.

The idea was to create an "aquarium-like" environment, dubbed the Vivarium, where artificial or simulated life forms could exist.

The whole Vivarium project was a significant research focus within the Media Lab, involving various teams working on different aspects.

Our specific project centered on Logo programming for the Apple Macintosh which I had developed at Terrapin - specifically the 512K model. This version of Logo incorporated features from TI Logos, like sprites, but with unique implementations including various procedures.

We decided to simulate squirrel behavior in our Vivarium. This decision stemmed from our observations of real squirrels' often perilous road-crossing antics. Our model comprised a simple road environment with a stoplight and intersection, where digital squirrels attempted to navigate the dangers of crossing a busy road. We incorporated moving cars and trucks, each programmed to mimic human-like responses at traffic lights - speeding up at yellow, sometimes running red lights, and so on.

The squirrels were modeled as Logo procedures, with each line representing a rule based on environmental factors like nearby cars, traffic light color, time, etc. These rules helped the squirrel decide whether to stay put or attempt crossing. A genetic algorithm aspect was introduced where, after successfully crossing several times, a squirrel's procedures would be combined with another successful squirrel's, simulating a form of digital reproduction.

We made no specific efforts to fine-tune the hyperparameters of our genetic algorithm; our focus was more on functionality. After a weekend of continuous running, we returned to observe a rather sobering yet enlightening outcome. A squirrel darted towards the road, hesitated at the sight of an oncoming truck, but eventually ran out and met its virtual demise. This outcome, while grim, was considered a success as it accurately mirrored real-world squirrel behavior.

To my knowledge, this project marked the first application of genetic algorithms within Logo programming to model squirrel behavior, a unique facet of the Vivarium project at the MIT Media Lab. This venture not only showcased the potential of blending genetic algorithms with traditional programming but also provided a fascinating insight into animal behavior through the lens of technology.