Solutions that Evolve

Building Self-Improving Systems

with Genetic Algorithms

Barry S. Stahl

Principal Engineer - AZNerds.net

@bsstahl@cognitiveinheritance.com

https://CognitiveInheritance.com

Favorite Physicists

1. Harold "Hal" Stahl
2. Carl Sagan
3. Richard Feynman
4. Marie Curie
5. Nikola Tesla
6. Albert Einstein
7. Neil Degrasse Tyson
8. Niels Bohr
9. Galileo Galilei
10. Michael Faraday

Other notables: Stephen Hawking, Edwin Hubble, Leonard Susskind, Christiaan Huygens

Favorite Mathematicians

1. Ada Lovelace
2. Alan Turing
3. Johannes Kepler
4. Rene Descartes
5. Isaac Newton
6. Emmy Noether
7. George Boole
8. Blaise Pascal
9. Johann Gauss
10. Grace Hopper

Other notables: Daphne Koller, Grady Booch, Leonardo Fibonacci, Evelyn Berezin, Benoit Mandelbrot

• See my presentation from CodeMash 2025: How the Fediverse Could Save Democracy & Why It Probably Won't

• Hear my thoughts on why a thriving Fediverse is so critical on David Giard's Technology and Friends show.

• Learn more about the Fediverse at fediverse.party.

• Get your own Mastodon account at Join Mastodon.

  • Don't stress on what server to use, just pick one and go. You can always move later.

• Follow me on Mastodon at @bsstahl@CognitiveInheritance.com.

Find the best strategy in a multi-player board game​

• Variant of Chutes & Ladders​
  • Player decides whether or not to take a chute or ladder​
• Extremely large Solution Space
  • Impossible to simply try all options

• Shortest Path Algorithm
  • Guarantees shortest path but not necessarily best strategy

• Mixed-Integer Programming Model
  • Requires a mathematical fitness function (can't use a simulation)

• Neural Network
  • Requires a very large training data corpus

Find Solutions by Simulating Darwinian Evolution​

• Each candidate solution is defined by its properties (chromosomes)​
• A fitness function is used to determine which solutions "survive"
  • A simulation may be used as the fitness function
• Surviving solutions may mutate and evolve other solutions​

• Optimal solution may never be found
  • i.e. If a optimality requires a large combination of elements
• Optimum may be found but not recognized
  • If conditions make that solution non-optimal at that moment
    • i.e. If a key feature is not used during the simulation
  • Minimize the liklihood by increasing simulation size

• We need a strategy that will tell us where we

  should move to next, given:

  • Our location on the board
  • The result of our spin
  • The location of our opponents (opt)
  • Our player number (opt)

• Example: If we are on square 45 and spin a:

  • 1 - 46 (no choice)
  • 2 - 47 (no choice)
  • 3 - 48 or 26
  • 4 - 49 or 27
  • 5 - 50, 11 or 28
  • 6 - 51, 12, 29 or 84

• A chromosome represents all of the possible decisions that can be made

• Each gene represents a decision

  • Ideally a gene fully describes a single game state
  • If no decision is possible for a state, there is no gene for that state

• An allele is an option for that decision

  • If there are 3 options for a given state there will be 3 possible alleles for that gene

• Simulations per Generation
  • Controls how often the solutions evolve
  • More simulations make it more likely the best solution will survive

• Generation Count
  • Controls simulation length
  • Optionally look for convergence
    • When does improvement stop?
    • When do significant changes stop during crossover?

• Error (misspelling) rate
  • Controls how frequently genes change allele's during the evolutionary process
  • Higher rates cause greater variations from generation to generation

• Selection strategy
  • Defines which solutions survive and mutate
    • Survive intact - Remain unchanged to next generation
    • Survive mutated - Modified versions survive to next generations
    • Survive recombined - Crossover between two solutions survives

• Implement in your preferred language

  • C# implementation is on GitHub

• Augment to include what place the player is in

  • i.e. Might make different decisions if started 1^st vs 6^th

• Augment to include opponent locations

  • i.e. Might make different decisions if opponents are close to the end

• Define the DNA of another game

  • Recommendation - stick with simpler games to start
  • Avoid games like chess with massive complexity

• A point in multi-dimensional space
• Mathematical representation of a word or phrase
• Encode both semantic and contextual information

• Model: text-embedding-ada-002
• Vectors normalized to unit length
• Use 1536 dimensions

Relate vectors based on the angle between them

• Cosine Similarity ranges from -1 to 1, where:

  • +1 indicates that the vectors represent similar semantics & context
  • 0 indicates that the vectors are orthogonal (no similarity)
  • -1 indicates that the vectors have opposing semantics & context

• Cosine Distance is defined as 1 - cosine similarity where:

  • 0 = Synonymous
  • 1 = Orthogonal
  • 2 = Antonymous

Note: For normalized vectors, cosine similarity is the same as the dot-product

Find distant embeddings

• Assumptions

  • No db of embeddings available to search
  • No understanding of features of the embedding space
    • i.e. changing language changes most dimensions

• What type of models might work?

  • Can we do it logically?
  • Can we brute-force it?
  • Are there hybrid approaches?

• What are the features of each model?

• How do we execute the model(s)?

• Categorical

  • Part of Speech (noun, verb, etc)
  • Register (formal, slang, technical, etc)
  • Morphology (root, compound, phrase)
  • Animacy (inanimate, animate, agentic, collaborative)
  • Scientific Discipline (biology, physical, life/health, etc)

• Continuous

  • Polarity (negative -> positive)
  • Idiomaticity (literal -> idiomatic)
  • Concreteness (abstract -> concrete)

• Other Options

  • Language

What should change to find words distant from the current

• Part of Speech: Noun -> Adverb
• Discipline: Software Engineering -> Philosophy
• Polarity: Neutral -> Highly Positive

You are a simulation of a great linguist.

You classify words/phrases within the following categories:

• Part of Speech (noun, verb, etc)
• Register (formal, slang, technical, etc)
• Scientific Discipline (biology, physical, life/health, etc)
• Morphology (root, compound, phrase)
• Animacy (inanimate, animate, agentic, collaborative)
• Polarity (highly negative, negative, neutral, positive, highly positive)
• Idiomaticity (highly literal, literal, idiomatic, highly idiomatic)
• Concreteness (highly abstract, abstract, concrete, highly concrete)

You identify words/phrases that meet the requested characteristics.

Give me a word or phrase with the following characteristics:

• Primary part of speech is verb
• Register is technical
• Morphology is phrase
• Animacy is collaborative
• Scientific Discipline is biology
• Polarity is highly negative
• Idiomaticity idiomatic
• Concreteness highly concrete

and is not in the following list: "prime genius", "engineer a cooperative symbiosis".

Be sure to only respond with the selected word or phrase, no ceremony.

Once we can encode the traits of a word/phrase/idea, what other fitness functions could we apply?

• The DNA defines the traits of the specific organism (candidate)
• The Fitness Function defines the direction of evolution

• Phonetics - Alliteration, Rhyme, or Assonance
• Rhythm - Punchy words or flowing patterns
• Imagery - Reward evocations of vivid emotional imagery
• Narrative - How “heroic”, “mysterious”, or “pastoral”
• Temporal - Prefer archaic, modern, or futuristic language
• Novelty - Reward surprise or pattern-breaking
• Constraint - Avoid forbidden categories or styles
• Information-Density - Pack more meaning per syllable

• We discussed 2 problems:
  • Game strategy - Best move given game state?
  • Semantic Distance - Word to test next?
• What do they have in common?
• What makes them candidates for genetic algorithms?

• Genetic Algorithms "Learn" by simulating Evolution
• The "DNA" represents the decisions that can be made
  • One chromosome per condition (game state)
  • One gene per option (choices from a state)
• Several parameters control how solutions evolve
  • Simulations per Generation - how many changes
    • Fewer simulations increases risk of less-fit surviving
  • Misspelling rate - how often chromosomes change
    • Higher rates cause greater variations
  • Selection strategy - Who survive/mutates
    • Survive intact - Remain unchanged
    • Survive mutated - Modified versions survive
    • Survive recombined - Crossover survives

• AI Demos Code Repo
• My Algorithm Articles
• My AI Articles

Attempts to make the best possible decision​ based on the model and the data

• Model - Our description of the problem
• Data - Our understanding of the state of the domain

The success of a problem-solving algorithm is often determined before the algorithm even runs -- by how the problem is modeled

• Modeling defines the problem space
  • What’s included & what’s ignored
  • Poor models lead to irrelevant solutions

• Modeling shapes the solution strategy
  • Logical models » OOP or Rules Engines
  • Graph models » traversal algorithms
  • Constraint models » LP or MIP
  • Probabilistic models » Neural/Bayesian Networks

• Modeling can determine tractability
  • Some formulations make problems NP-hard
  • Others reveal polynomial-time solutions
  • Examples
    • Conference Scheduling with LP vs MIP
    • Traveling Salesman with OOP