Directing the evolution
By directing the evolution we’re augmenting our creativity — without the need to acquire new knowledge.
[IN BRIEF] This is an interpretation of how recent advances are surpassing the scientific method, taking advantage of the generative force of chaos instead of relying uniquely on the human mind and rational explanations. The new emerging method extends the human ability to create new things, therefore going far beyond the limits of the theories. My effort is to give an unifying perspective, highlighting common traits that are emerging in different application fields, including generative design, directed evolution, and artificial intelligence. I herein underline the need of a unifying philosophy or framework that could organize the “evolutionary method” independently from its application field. Human creativity is going to be augmented: new amazing things are going to populate our world, and these things aren’t coming uniquely from human mind…
Creating with knowledge
If you want to create something new, you have to know how to do it.
· Behind the creation of a structure (e.g.: a skyscraper), there are the laws of statics
· Behind the creation of a new enzyme-based chemical product (e.g.: a detergent), there are the laws of chemistry
· Behind the creation of a new information processing system (e.g.: computer program), there the underlying processes or algorithms.
· And so on…
Knowing the theory is traditionally essential for rationally creating new things. The theory is a synthesis that generalize one or more phenomena. In the general laws, we can find the solution to particular problems.
When you create something new, you can clearly explain each component of the system: why it is there, and how it works.
Creating without knowledge
But what if we don’t know the theory? we navigate in the dark forest of ignorance…
Having no theory: is the kind of situation when you are in front of your True-or-False examination and you studied nothing; you’re totally not prepared. So you do some random guessing: T,F,F,T,F,T….. random guessing is the opposite of a rational approach, but, after all, in this case, is your only choice.
Doing random guessing in the real world is absolutely not desirable. I wouldn’t appreciate if the engineer who designs my new house uses the lottery numbers to size the main pillar carrier. If you want to create something new in the real world, you must have a solid background.
But there are problems for which there is a real lack of theory. Problems that involve very complex structures. Or contain million or even billion of free parameters. Problems dealing with something mysterious or only partially explained. For example:
· we would like to design a structure constrained by very demanding constraints and we don’t know how to manage its highly interconnected complexity
· we would like to synthesize a certain protein with certain desirable features, but we don’t know how to implement a process that could generate it
· we would like to realize a new form of intelligence but can’t codify an explicit procedure, either because is very complex or because we can’t completely understand it explicitly.
The generative force of evolution
Before going on, I would like to spend a couple of words on the huge generative force of evolution.
Evolution is a sort of magic process that produces order from disorder. All what is around us - and even us — are the result of an evolutionary process. It’s clear that evolution has an incredible creative force.
It may seem a trivial statement but all what has been created, has been created without a theory. Evolution does know no theory. And, without any theory, the creative force of evolution gave amazing things, including:
· Complex Structures — like the giant sequoias
· Chemical processes and reactions — like the photosynthesis
· Forms of intelligence — like the brains: from smaller (ants) to bigger (human)
The intimate secret of evolution is time +freedom. Freedom to get it right, freedom to do mistakes. Freedom to mutate, freedom to change the rules or overwrite them. “Trial and error” is commonly seen as a key element of personal growth, and a key element of all the entrepreneurial cultures. And even the scientific method brings with itself a component of try and re-try. At the very beginning of the Holy Bible, Genesis 1:22, God says to the man and the woman: “You may freely” — thus recognizing freedom as the primary source of growth and development. The randomness of the world gives to the creatures the capability to mutate, change, evolve and improve.
Directing the evolution: the general form
Let’s now see how we can empower ourselves with the incredible force of evolution to generate new amazing things.
If the nature could create amazing things without a theory, so we can. As we’ve seen, we “just” need time and freedom. Therefore, we need basically 2 things:
1. A FAST “MICROWORLD”: let’s imagine to have a little time bubble, where things happen very *very* fast. This is the enabling factor of the new evolutionary tecnique: an isolated fast micro-world. Very bad things could happen inside it without any direct consequence on the real world. The environment, includes a good part and hostile part; and also rules (that partially limit the freedom of the agents).
2. FREE AGENTS Are the second component. Could be only one type of agent, trying to survive in the static environment. Or multiple agents that collaborate. Or even more fascinating, two or more agents fighting each other, constantly pushing up their capabilities.
O.K., and what about the direction? “Direct” comes from latin dirigere, that means govern, give a sense. Depending on how we initially build (or select) the agents, and how we initially build the environment we somehow direct the future evolution of the agents — and we somehow “persuade” them to reach our final goal.
Proposed general form: “Directing the Evolution is the a something finding process that starts with user-defined goals and then induce some agents to evolve to explore multiple possibilities and select (and optionally amplify) the desired something. Could be implemented with fast computation systems (like computers, DNA/RNA, or non-conventional c.d. like molds)”.
Directing the evolution: practical applications
Generative Design: It’s a form finding process that starts with user-defined design goals and then explores innumerable possible solutions to select the best design option. It’s typically computer-implemented — virtual agents.
Directed evolution: It’s protein finding process that starts with user defined goals and then subjecting a gene to iterative rounds of mutagenesis, selection and amplification to produce the desired protein. It’s typically in-vitro implemented — biological agents.
Training Neural Networks: It’s an intelligence finding process that starts with user defined goals and then induce a neural network architecture to get the weights that give it the desired form of intelligence. It’s typically computer-implemented — virtual agents.
Volta Robots: a self-driving rover in forest trails
My Company, Volta Robots, is specialized in developing forms of intelligence for small, unmanned vehicles. This year we’ve been exhibitors at CES Las Vegas, and we unveiled a self-driving rover that can autonomously follow forest trails, relying only on visual input (no GPS, no lidars, etc…).
We achieved this impressive, self-improving rover by implementing an evolutionary algorithm on a relatively large CNN. The rover can find its way in a wide range of terrains, demonstrating that has internalized a very good generalization of what a “trail” is — and can learn from its errors, that become less and less frequent.
The following neural network with a 5-pixels input layer (5 distance points from borders of the street), two hidden layers (7 neurons on two levels) and two output neurons (steer and thrust). The network has been trained genetically: each generation produces an high number of “bad” cars, but, among them, there’s one carrying good neural network weights and travels more then others.
What we’ve done with our Volta Rover is essentially a scale-up, with 640x320 px multi frame input with LSTM. However the evolutionary logic is exactly the same. The resulting Volta Rover can travel for several miles in total autonomy, exploring different unseen scenarios and trails without any sensor, relying only on one front-facing camera. This is an example of how we’ve directed the evolution of something we don’t know. We can’t explicitly code a computer program that do the same simply because we don’t know the underlying procedure.
Some really inspiring papers moving further in the evolutionary (=> self-improving, somehow “unsupervised”) sense:
- Generative Adversarial Nets
Ian Goodfellow et al.—
- Evolution Strategies as a Scalable Alternative to Reinforcement Learning
Tim Salimans, Jonathan Ho, Xi Chen, Ilya Sutskever et al. — https://arxiv.org/pdf/1703.03864.pdf
- Natural Evolution Strategies
Jurgen Schmidhuber, Daan Wierstra, Tom Schaul et al. — http://www.jmlr.org/papers/volume15/wierstra14a/wierstra14a.pdf
- https://medium.com/@peterbinggeser/designing-ai-solving-snake-with-evolution-f3dd6a9da867 a simple example of how to model and evolve a snake in the “snake game”.
- http://www.evolvingai.org/robotics this is an amazing example of robot adaptation
Caltech: a carbon-silicon protein
Recently Fances Arnold, a pioneer in the directed evolution, “convinced” bacteria to produce selectively make silicon-carbon bonds 15 times more efficiently than the best catalyst invented by chemists.
The used technique enlists the help of nature’s design process — evolution — to come up with better enzymes — “We can do what nature takes millions of years to do in a matter of weeks,” says Arnold.
In a micro-world, an initial condition has been created by selecting a protein from a bacterium that grows in hot springs in Iceland, that somehow they know could evolve into their goal. Then iteratively:
- they mutate the DNA coding for that protein within a region that specifies an iron-containing portion of the protein thought to be responsible for its silicon-carbon bond-forming activity. The DNA coding for the enzyme is mutated in more-or-less random ways.
- They test these mutant enzymes for their ability to make organosilicon compounds better than the original (desired traits => “direction”).
- The top-performing enzyme is then mutated again, and the process is repeated until an enzyme that performs much better than the original is created.
After only three rounds, they had created an enzyme that can selectively make silicon-carbon bonds. The enzyme is highly selective, which means that it makes fewer unwanted byproducts that have to be chemically separated out.
Autodesk: a drone frame
On the CAD giant’s webpage dedicated to generative design we read: “Software mimics nature’s approach to design.” and “Quickly generate high-performing design alternatives — many that you’d never think of on your own — from a single idea.” In 23 words they clearly highlight the radical change that’s happening in human creativity. The project starts with a blurry idea, a dream. Then, this creative inspiration take multiple explorable forms, with the help of a computer-implemented evolutionary algorithm. These 3 steps could be repeated multiple times, until the “dream has been catched”.
In this self-explicative video Autodesk demonstrates how to create a drone frames (and other objects) with generative design (Dreamcatcher).
Implication of the Evolutionary Method
Wrapping-up: so far we’ve seen that a new “Evolutionary Method” is going to augment our creativity. New performing things could be obtained exploiting the same generative force that created forms of life: the evolution. With this new superpower we’re augmenting our capability to create new forms of intelligence, new structures and new forms of life. We’ve also seen that this method is different from the scientific method.
Evolutionary vs. Scientific Method — the role of chaos Poor Galileo didn’t have an isolated and fast micro-world. And building up a cathedral by trial and errors in the real world was not so convenient: he needed a theory and a rational approach that could predict the implication of each design choice. But with the capability to create fast, isolated micro-worlds and populate them with free agents: everything is changing.
The first implication is: Chaos, traditionally considered negative — the natural enemy of theories and rational explanations — now becomes a developing factor.
The need of an unifying framework — As we’ve seen there are at least three different fields already empowered by directed evolution and sharing common traits: in structure design (with generative design), in chemistry (with directed evolution) and in artificial intelligence (with the training of neural network). “Evolution” has specific meaning in each field — and I apologize if I’ve been unprecise in the specific use of the term “evolution”. However, my goal here is to highlight a trend common to different fields. We should create a (super-)theory on how to direct the evolution, this means a set of techniques and knowledge about how to create something unkwnown though an evolutionary process.
The second implication is: we need a unified phylosophy or general framework describing the general form of the “Evolutionary Method”. This should be cross-disciplinary and agnostic of the implementation method (computer, bacteria, other…). A particular enphasis should be put on how to manage the “reproducibility issue” and document results in the scientific community.
Very complex objects are going to populate our real world, and these objects are coming directly from “directed micro-worlds”. These things are something “alien”, in the sense that doesn’t come nor from the evolution of our world, nor directly from human mind. They came from directed evolution happened inside artificially directed micro-worlds. We are already using alien things such as structures, proteins, e forms of intelligence, for which we’ve directed the evolution and are not rationally designed. These upcoming things are like dreams: we don’t have a full conscious understanding of their inner workings — but, actually, they work.
The place couldn’t be better: not only for the amazing italian villa with beautiful gardens, not only for the sunny warm day in the middle of the pre-alps, and not only for the superior quality of the food (and beverages including Rossi d’Angera liquors and Angelo Poretti beers). In-fact, inside the villa, there is a huge wall painting (of course original) of Giuseppe Bertini, imaging Galileo Galilei (1564−1642) showing the telescope to the Doge of Venice (Venezia). And, as you know, the scientific method was fashioned mostly entirely by Galileo Galilei. There’s no better place to talk about a the new evolutionary method to create new things. In fact, it doesn’t require a complete model of phenomena, and — in this sense — goes far beyond the scientific method.