Title: Algorithmic Insights and the Theory of Evolution
1Algorithmic Insightsand the Theory of Evolution
- Christos H. Papadimitriou
- UC Berkeley
2The Algorithm as a Lens
- It started with Alan Turing, 60 years ago
- Algorithmic thinking as a novel and productive
way for understanding and transforming the
Sciences - Mathematics, Statistical Physics, Quantum
Physics, Economics and Social Sciences - This talk Evolution
3Evolution Before Darwin
4Before Darwin
5Before Darwin
- Charles Babbage
- Paraphrased
- God created not species, but the Algorithm for
creating species
6Darwin, 1858
- Common Ancestry
- Natural Selection
7The Origin of Species
- Possibly the worlds most masterfully compelling
scientific argument - The six editions 1859, 1860, 1861, 1866, 1869,
1872
8The Wallace-Darwin papers
9Brilliant argument, and yet many questions left
unasked, e.g.
- How does novelty arise?
- What is the role of sex?
10After Darwin
- A. Weismann
- Paraphrased
- The mapping from genotype to phenotype is
one-way
11Genetics
- Gregor Mendel 1866
- Number of citations
- between 1866 and 1901
- 3
12The crisis in Evolution1900 - 1920
- Mendelians vs. Darwinians
- Geneticists vs. Biometricists/Gradualists
13The Modern Synthesis1920 - 1950
Fisher Wright - Haldane
14Big questions remaine.g.
- How does novelty arise?
- What is the role of sex?
15Evolution and Algorithmic Insights
- How do you find a
3-billion long string in 3 billion years?
- L. G.Valiant
- At the Wistar conference (1967), Schutzenberger
asked virtually the same question
16Valiants Theory of Evolvability
- Which traits can evolve?
- Evolvability is a special case of statistical
learning
17Evolution and CS PracticeGenetic Algorithms
ca. 1980s
- To solve an optimization problem
- create a population of solutions/genotypes
- who procreate through sex/genotype
recombination - with success proportional to their objective
function value - Eventually, some very good solutions are bound
to arise in the soup
18And in this CornerSimulated Annealing
- Inspired by asexual reproduction
- Mutations are adopted with probability increasing
with fitness/objective differential - (and decreasing with time)
19The Mystery of Sex Deepens
- Simulated annealing (asexual reproduction) works
fine - Genetic algorithms (sexual reproduction) dont
work - In Nature, the opposite happens Sex is
successful and ubiquitous
20?
21A Radical Thought
- What if sex is a mediocre optimizer of fitness (
expectation of offspring)? - What if sex optimizes something else?
- And what if this something else is its raison
d être?
22Mixability!
- In LPDF 2008 we establish through simulations
that - Natural selection under asex optimizes fitness
- But under sex it optimizes mixability
- The ability of alleles (gene variants) to perform
well with a broad spectrum of other alleles
23Explaining Mixability
- Fitness landscape of a 2-gene organism
Entries fitness of the combination
Rows alleles of gene A
Columns alleles of gene B
24Explaining Mixability (cont)
- Asex will select the largest numbers
25Explaining Mixability (cont)
- But sex will select the rows and columns with the
largest average
26Neutral Theory and Weak Selection
- Kimura 1970 Evolution proceeds not by leaps
upwards, but mostly horizontally, through
statistical drift - Weak selection the values in the fitness matrix
are very close, say in 1 e, 1 e
27Changing the subjectThe experts problem
- Every day you must choose one of n experts
- The advice of expert i on day t results in a gain
Gi, t in -1, 1 - Challenge Do as well as the best expert in
retrospect - Surprise It can be done!
- Hannan 1958, Cover 1980, Winnow, Boosting,
no-regret learning, MWUA,
28Multiplicative weights update
- Initially, assign all experts same probability
- At each step, increase the probablity of each by
(1 e GI, t) (and then normalize) - Theorem Does as well as the best expert
- MWUA solves zero-sum games, linear programming,
convex programming, network congestion,
29Disbelief
- Computer scientists find it hard to believe that
such a crude technique solves all these
sophisticated problems - The eye to this day gives
- me a cold shudder.
- cf Valiant on three billion bits and years
30- Theorem CLPV 2012 Under weak selection,
evolution is a game - the players are the genes
- the strategies are the alleles
- the common utility is the fitness of the organism
(coordination game) - the probabilities are the allele frequencies
- game is played through multiplicative updates
31There is more
- Recall the update (1 e Gi, t)
- e is the selection strength
- (1 e Gi, t) is the alleles mixability!
- Variance preservation multiplicative updates is
known to enhance entropy - Two mysteries united
- This is the role of sex in Evolution
32Pointer Dogs
33Pointer Dogs
C. H. Waddington
34Waddingtons Experiment (1952)
Generation 1 Temp 20o C
35Waddingtons Experiment (1952)
Generation 2-4 Temp 40o C 15 changed Select
and breed those
36Waddingtons Experiment (1952)
Generation 5 Temp 40o C 60 changed Select
and breed those
37Waddingtons Experiment (1952)
Generation 6 Temp 40o C 63 changed Select
and breed those
38Waddingtons Experiment (1952)
() Generation 20 Temp 40o C 99
changed
39Surprise!
Generation 20 Temp 20o C 25 stay changed!!
40Genetic Assimilation
- Adaptations to the environment become genetic!
41Is There a Genetic Explanation?
- Function f ( x, h ) with these properties
- Initially, Prob x p0 f ( x, h 0) 0
- Then Probp0f ( x, 1) 15
- After breeding Probp1f ( x, 1) 60
- Successive breedings, Probp20f ( x,1) 99
- Finally, Probp20f ( x, 0) 25
42A Genetic Explanation
- Suppose that red head is this Boolean function
of 10 genes and high temperature - red head x1 x2 x10 3h 10
- Suppose also that the genes are independent
random variables, with pi initially half, say.
43A Genetic Explanation (cont.)
- In the beginning, no fly is red (the probability
of being red is 2-n) - With the help of h 1, a few become red
- If you select them and breed them, 60 will be
red!
44Why 60?
45A Genetic Explanation (cont.)
- Eventually, the population will be very biased
towards xi 1 (the pis are close to 1) - And so, a few flies will have all xi 1 for all
i, and they will stay red when h becomes 0
46Generalize!
- Let B is any Boolean function
- n variables x1 x2 xn (no h)
- Independent, with probabilities
- p (p1 p2 pn)
- Satisfiability game if B is satisfied, each
variable gets 1, otherwise 1 - e - Repeated play by multiplicative weights
47Boolean functions (cont.)
- Conjecture This solves SAT
- Can prove it for monotone functions (in poly
time) - Can almost prove it in general
- (Joint work with Adi Livnat, Greg Valiant, Andrew
Won)
48Interpretation
- If there is a Boolean combination of a modestly
large number of genes that creates an
unanticipated trait conferring even a small
advantage, then this combination will be
discovered and eventually fixed in the
population. - With sex, all moderate-sized Boolean functions
are evolvable.
49Sooooo
- The theory of life is deep and fascinating
- Insights of an algorithmic nature can help make
progress - Evolution is a coordination game between genes
played via multiplicative updates - Novel viewpoint that helps understand the central
role of sex in Evolution
50Thank You!