The creative evolution of digital organisms

Experimental evolution at Montpellier 

On the 22nd of August (2018), I had the pleasure to present work on Virtual Microbes at II Joint Congress on Evolutionary Biology. It was very inspiring to finally meet some of the people who designed/conducted the awesome experimental evolution studies I came to know and love over the past 5-10 years. It was also somewhat a relief to see how, vice versa, Virtual Microbes were well received by said experimentalists. Before this becomes any more self-serving, I'd just like to share this tweet by Vaughn Cooper: "Evolved populations learn to trust us, but die if not fed". I could not have put it any better :)

My presentation however, was a bit of an odd one out. Most of the work presented in the session was (quite understandably) about "real" experiments, such as the 30-year evolution experiment on Escherichia coli (@RELenski), using lineage tracing to track fitness relevant traits (@GrantKinsler), using media-in-oil emulsions to measure the trade-off between growth rate and yield (@katrinavanraay), and determining the energetic consequences of artificially selecting algae for body size (@MartinoMalerba). With all these awesome new experimental techniques, why do we bother simulating evolution in our computers? This question was answered by a presentation earlier that morning by Dusan Misevic (@dulefr), entitled "The surprising creativity of digital evolution". The work presented here, illustrated evolution's astonishing ability to do something quite interesting: debug our intuitions[efn_note] Lehman, Joel, et al. "The surprising creativity of digital evolution: A collection of anecdotes from the evolutionary computation and artificial life research communities." arXiv preprint arXiv:1803.03453 (2018). [/efn_note]. However, many of the "surprises" in this work are best visualised in animations, so I decided that a blog was just the place to summarise some of the insights, and show a few other cool examples of digital evolution.

The Surprising Creativity Evolution

Like all science, evolutionary biology is in the business of surprising us. Great examples from evolutionary biology are the reprogramming of host brains by parasites[efn_note]Lefevre T, Adamo SA, Biron DG, Misse D, Hughes D, Thomas F. Invasion of the body snatchers: the diversity and evolution of manipulative strategies in host–parasite interactions. Advances in Parasitology. 2009;68:45–83.[/efn_note] [efn_note]Maeda, Hiroki, et al. "Positive phototropism is accelerated in Biomphalaria glabrata snails by infection with Schistosoma mansoni." Parasitology international 67.5 (2018): 609-611.[/efn_note], the evolution of cooperation that can occur especially when it is costly [efn_note]Colizzi, Enrico Sandro, and Paulien Hogeweg. "High cost enhances cooperation through the interplay between evolution and self-organisation." BMC evolutionary biology 16.1 (2016): 31.[/efn_note], and the pebble toad's weird strategy of getting away from scary spiders[efn_note]Pebble toad rollover - Nature's Greatest Dancers, preview episode 2. https://youtu.be/yrw-6KguB8E[/efn_note].

(BBC - Pebble toad rollover - Nature's Greatest Dancers)

Surprises in Digital Evolution

Digital evolution is no different, and has resulted in a collection of awesome ancedotes over the last couple of decades. I here discuss my favourites from the article by Joel Lehman et al. (2018).

Why walk when you can somersault?

Work by: Karl Sims (1994)
Objective: getting blocks (controlled by a neural network) to evolve locomotion
Fitness: Velocity

The evolved creatures evolve to exploit a bug in the physics engine, using the initial potential energy to flip in just the right way. Note that eventually, this exploit of the physics engine was resolved and endless forms of locomotion did evolve

https://youtu.be/TaXUZfwACVE?list=PL5278ezwmoxQODgYB0hWnC0-Ob09GZGe2


Creative falling

Work by: Peter Krcah (2008)
Objective: getting blocks (controlled by a neural network) to evolve jumping
Fitness: Distance from the ground (w.r.t. the block that was originally closest to the ground)

Although not necessarily exploiting a bug (like before), the organisms now exploit the fact that the researchers did not define their fitness criterium precisely enough. Instead of jumping, it flips its legs to move the block closest to the ground up!

https://youtu.be/N9DLEiakkEs?list=PL5278ezwmoxQODgYB0hWnC0-Ob09GZGe2

Elbow walking

Work by: Antoine Cully (2015)
Objective: getting robots to walk with a broken leg
Fitness: Walking without one (or more) legs touching the ground

The robots quickly learned to avoid walking on the legs it was not allowed to. Surprisingly, when 0 legs were allowed to walk on, evolution found a way!

https://youtu.be/H6OB1E8NsLw?list=PL5278ezwmoxQODgYB0hWnC0-Ob09GZGe2&t=114

Spinning towards the light

Work by: Richard Watson and Sevan Ficici (2002)
Objective: getting driving robots to move towards a light source
Fitness: Amount of time to reach the light source

Although the Braitenberg behaviour (left) is actually a very quick way to move towards the light, the robots most often evolved spinning behaviour (right). This spinning behaviour turned out to be much easier to "find" by evolution.

What do we learn from these anecdotes?

In reviewing the creativity of digital evolution (Joel Lehman et al., 2018), we have unveiled how evolution can help us "debug" the following things:

  • What we think we are selecting for isn't necessarily the same as what is actually selected for
  • Our experiments and simulations are vulnerable to exploits
  • We generally underestimate the power of evolution
  • Convergence with real biology reveals that certain solutions might be "easier" than others

Digital Evolution at its best

While simulated evolution has yielded all the amusing anecdotes shown above, it actually works very well most of the time. See the video's below for some cool examples:

https://youtu.be/z9ptOeByLA4?list=PL5278ezwmoxQODgYB0hWnC0-Ob09GZGe2

Unshackling Evolution: Evolving Soft Robots with Multiple Materials and a Powerful Generative Encoding. Cheney, MacCurdy, Clune, & Lipson. Proceedings of the Genetic and Evolutionary Computation Conference. 2013.

https://youtu.be/pgaEE27nsQw

Geijtenbeek, Thomas, Michiel Van De Panne, and A. Frank Van Der Stappen. "Flexible muscle-based locomotion for bipedal creatures." ACM Transactions on Graphics (TOG) 32.6 (2013): 206.

https://youtu.be/Aut32pR5PQA
A small 2D simulation in which cars learn to maneuver through a course by themselves, using a neural network and evolutionary algorithms. (Samuel Arzt on Youtube)

Evolution Simulator

To wrap up, I'd like to share this online "game" called Evolution Simulator which you can play online or download for free! Instead of writing an actual finishing statement for this blog, I'll just go ahead and play this game for a bit. ;)

References

Previous
Previous

Glucose metabolism in Virtual Microbes