What follow is a very slightly cleaned up version of the notes I took during the first ISAL summer school session at ECAL 2017. Note that this is very rough, I was just reviewing my notes and thought I may as well put them into a post format at the same time to revisit later. Apologies for any typos or other mistakes.
Historical and Philosophical Perspectives on Artificial Life
Mark A. Bedau, Reed College
Artificial life enables us to learn fundamental properties of living systems by synthesizing them in artificial media
Kinds of alife:
- “soft” - synthesize in software
- “wet” - wetware (biochemistry/wet lab)
- “hard” - hardware, physical systems
- “mixed” - social-technical systems
Why life is special:
- mysterious - entices fundamental questions we don’t understand yet
Fundamental properties of life:
- adaptive complexity
“Life deserves our scrutiny more than you might think if you just take some biology classes”
Birth of artificial life:
- 1989 - 1st Artificial Life conference - Chris Langton, study of “life-as-it-could-be”
Historical roots: Reproduce and study key capacity of living systems in simple formal models.
Before the above conference, others made importnat advances the field grew out of; they are the “founding grandfathers” of the field:
- John von Neuman - cellular automata
- Norbert Wiener - cybernetics
- John Conway - game of life
- John Holland - genetic algorithms: practical software tool which will work and build useful systems by an artificial selection-like process
Artificial life and complex systems
The field of artificial life today is connected with the study of complex systems.
- complex adaptive systems
- many locally interacting elements. Large things happen through small interactions (simple rules, large global effect)
- simple rules governing elements can/do change over time
It is extremely hard to understand complex systems by analyzing them; behaviour is dependent on interactions. The main practical way to study this is to build it, then step back and observe.
- synthesis and simulation
- “What I cannot create I do not understand” - Richard Feynman
A grand challenge in artificial life
There appears to be a trend with all life - qualitative complexity. Complex organisms become more complex over time. Example:
- prebiotic soup ->
- simple prokaryotic cells ->
- complex eukaryotic cells ->
- simple multi cellura life ->
- large bodied vertebrates ->
- sophisticated sensory capacities ->
- sophisticated communication, language
As time goes on, complexity in life has increased. How? Why? What explains the above trend?
Deep law of nature?
When organized energy streams down onto earth, much of it simply turns into disorder. However, something else also happens…there are processes of organization rather than disorganization.
- Farmer & Belin in Artificial Life II (1992) [see: second law of thermodynamics]
Thus, from the war of nature, from famine and death, the most exalted object which we are capable of conceiving, namely, the production of higher animals, directly follows. There is grandeur in this view of life, with its several powers, being originally breathed into a few forms or into one; and that, whilst this planet has gone cycling on according to the fixed laws of gravity, from so simple a beginning endless forms most beautiful and most wonderful have been, and are being evolved.
- Darwin, Origin of Species [see: natural selection]
So haven’t we already explained this?! I learned about natural selection in middle school!
Before we start, some terms:
- “trend” - empirical fact
- “robust regularity” - general trend; maybe some exceptions, not accidental but typical
- “mechanism” - how a trend is produced - accidental or robust.
Arrow of complexity hypothesis
Evolution inherently creates increasingly complex adaptive organizations
The arrow of complexity is a robust regularity. It has constraints and exceptions, and is not accidental. No mechanism is specified.
Testing the arrow of complexity hypothesis
According to Stephen Jay Gould in “Wonderful Life”, all we have to do is “replay the tape of life”:
- rewind and erase
- replay forward
- new contingencies
- repeat again and again
- see robust regularities
This would be an excellent test if we could actually do it.
Gould’s final analysis:
- The increased complexity is completely accidental;
- Contingent historical process, thus no laws
- “Almost every interesting event of life’s history falls into the realm of contingency…any replay of the tape would lead evolution down a radically different pathway from the road actually taken”
He concluded that there is no reason to think things would get more and more complex because no laws govern contingent processes.
The speaker’s diagnosis is that Gould did not actually replay the tape of life - he just thought about it. The second law of thermodynamics is highly contingent in itself and yet there is a law. There is a difference between imagining replaying the tape and actually replaying the tape
Dan Dennett’s analysis - Darwin’s Dangerous Idea
- adaptive complexity is a “forced move” (as in chess)
- confirms and explains arrow hypotehsis
forced move: one obviously really good move; you are not physically forced to make it, but are “forced” by good chess playing because it is obviously the best move.
Problems with this:
- Is the move of increasing complexity “really” forced? It’s not clear that intelligence is a forced move; it has costs, which could outweight benefits. Eg bacteria are relatively unintelligent, yet they are thriving.
- Is there enough force? Even if intelligence is a good idea, is it a good enough idea? Eg having x-ray eyes may be good, but we aren’t evolving x-ray eyes all over the place (yet?)
Dennett thinks evolution is always going to create something that is more complex because it is a “forced move” (eg intelligence)
Speaker’s diagnosis: Dennett did not replay the tape, either. Like Gould, he just sat and thought about it a whole lot.
So - most people are just sitting in their office imagining replaying the tape, they are not replaying the tape.
How to actually replay the tape
- study natural analogs - eg island in the pacific. But problems: how did life forms get to the island? Where did they come from? They’re just colonized from existing life forms…not a true replay
- if simple, solve mathematical models. but the systems are not simple…
- if complex, simulate models with a computer - no magic or wishful thinking. The great thing is you may have theories in what may happen, but what simulation actually shows may be different. Your predictions won’t influence the outcome [note: a couple of the talks I saw later actually made me question this]
“Put your model where your mouth is”
Gould and Dennett told a nice story, but they did not try to independently verify it. Without such a model, it is hard to see how you will get past verbal wars and stories.
This model has species with some complexity, and there is some process by which once in a while a species will give birth to a new species and go extinct.
Life starts off being very basic. Randomly choose if new species is more or less complex than the previous species. There is some chance of species going extinct.
Whether a species becomes less or more complex is random - there is no assumption of a forced move.
Results of the model: complexity does increase. The model vindicates the arrow of complexity hypothesis
Issues with this model:
- produces merely nominal “complexity”
- nothing in the model is REALLY complex. just a data structure with a “complexity” value;
- assumes accessible space of unlimited complexity
- how is this space constructed and structured? eg can we REALLY get to xray eyes?
- how does evolution change and enlarge it?
- the space does all the work.
How NOT to beg the hard questions
What we want in a model:
- unlimited number of genetic possibilities
- natural selection
- endogenous fitness and co-evolution
- evolution constructs the enviornment
- computational universality
Tierra is close to the above. It takes computer memory & puts in a program created by hand that copies itself; one ancestor at gen0, the rest of the memory is empty. The program runs for a while and eventually copies itself. Eventually whole memory fills up.
- if memory gets full, you randomly kill some of the programs to create new space.
- allow mutations to happen. Once in a while an instruction is changed.
- most of the time bad instructions just crash, but sometimes they don’t. Once in a while they are actually better/reproduce faster.
Some programs developed into “parasites”. They were shorter than others and could not copy themselves on their own. But if they were next to another organism that does have copy instructions, they could use that organism to copy themselves. So they’ve thrown away many of their instructions because they do not need them. They end up reproducing faster because they are less complex. When these parasites start becoming too successful, they actually begin dying out as there end up being fewer hosts with copy functions to take advantage of. And so the model goes back and forth between host and parasite dominance.
Tierra complexity is NOT just nominal
- had both simple and complex programs, actually complex vs just having a complexity value of some sort [note, is the size of the program really a good representation of complexity here? What if the program is just filled with the same instruction over and over?]
- emerging space of evolutionary possibilities (what’s possible to evolve depends on what is already there)
- the space has real structure
- evolution changes the space’s structure.
But does this confirm the arrow of complexity hypotehsis?
Lesson from Tierra:
- evolution in Tierra is not very creative
- Tierra shows these mechanisms are NOT sufficient to explain the arrow
- most complex creatures in Tierra do not always get more complex…things often get simpler
No - Tierra didn’t quite confirm the arrow of complexity hypothesis.
The following are not by themselves sufficient to produce the arrow. Variations of them may be, but not simply their existence:
- natural selection
- endogenous fitness and coevolution
- construction of environment
- infinite genetic space
- computational universality
Darwin thought he proved the arrow via natural selection, but in the above model it’s not true! Natural selection is clearly not sufficient.
This remains one of the grand challenges in artificial life.