- tags
- Evolution, Complexity, Artificial Intelligence
- resources
- Open-endedness: The last grand challenge you’ve never heard of
Objectives and open-endedness
Objective functions are commonly used in all areas of machine learning (even so-called unsupervised learning, or evolutionary strategies). However this seems to be fundamentally opposed to how natural evolution proceeds. Innovations appear without a priori objectives and it isn’t clear if setting such objectives for ourselves is slowing down progress by putting the focus on too narrow paths to a solution (Woolley, Stanley 2011; Stanley, Lehman 2015).
My idea of Open Ended evolution thoughtsopen-ended-evolution
As part of my research, I have been thinking a lot about open-ended evolution and what this concept means to me. Although it is still early in my research journey, I will go into the process of writing down my thoughts about this very challenging concept in order to have material on which I will be able to reflect later and understand how my beliefs have changed.
I see open-ended evolution (OEE) as the holy grail of Artificial life (and maybe Artificial Intelligence) research. The only manifestation of intelligence we know of has emerged through a long and complex process, which inner workings we seem to understand. This process is often called evolution and sometimes reduced to one of its components: Natural Selection. Evolution obviously encompasses much more, starting from the process through which new characters and behaviors can emerge.
When people from the computer science community refer to OEE, they usually mean artificial OEE. That is, the creation of artificial systems (usually computer simulations?) that can exhibit the same properties as natural evolution. The main problem is: What are those properties?. I know many researchers have spent a lot of time figuring this out, and this part of the problem is certainly the most important one. As Hiroki Sayama once told me, citing another Alife researcher: Give me a definition of OEE and I can build it. Here are some properties I would expect from an open-ended system from the top of my head:
- Ability to grow in complexity. From a given state of the system, we can expect it to grow in complexity by building on previous complexity and adding innovations. This is probably the biggest open question, because although we can easily think about systems that keep generating novel things, we don’t know if they can grow in complexity. And even for things where there is evidence that it is possible, we don’t know how efficient they are.
- Adaptability to new environmental conditions. The system should have some amount of plasticity, or capability to change its internal functions in response to a change in external conditions. It is hard to tell if this is just a spontaneous property of systems satisfying the condition above or something more?
- Ability to communicate information to the outside of the system (this is not required for any OEE system per se but necessary if we want to be able to get something out of the simulation). One could also argue that complexity means nothing without an observer.
This list is certainly not complete, and probably not correct. But I believe these objectives are already hard enough to make the scientific challenge they represent exciting. I believe the cellular automata I am studying have the potential to qualify for the first property and the third. However, it is still not clear how efficient this can be and if it is within reach.
Interesting idea about open-endedness open-ended-evolution
Interesting idea presented in a paper by Sayama and Pattee (Pattee, Sayama 2019).
Bibliography
- Brian G. Woolley, Kenneth O. Stanley. . "On the Deleterious Effects of a Priori Objectives on Evolution and Representation". In Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation, 957–64. GECCO '11. New York, NY, USA: Association for Computing Machinery. DOI.
- Kenneth O. Stanley, Joel Lehman. . Why Greatness Cannot Be Planned: The Myth of the Objective. Cham, Switzerland: Springer International Publishing.
- Howard H. Pattee, Hiroki Sayama. . "Evolved Open-endedness, Not Open-ended Evolution". Artificial Life 25 (1):4–8. DOI. See notes