brief sketch for answer-factory manual design spike

20160128-answer-factory.md

1. build a relatively large dataset (1000+ training cases)
2. wire up postgres store

• answers
  • id (key)
  • program (index)
  • genome (index)
  • instance-count
  • scores
  • timestamp

3. 100 random answers -> datastore
4. score all answers
5. apply lexicase selection to get 100 parents, produce 200 children by crossover -> datastore
6. score all answers
7. apply lexicase selection to get 100 parents, produce 100 mutations with close ERC values -> ddatastore
8. repeat from 4

• no death
• monitor scores (including total-error)
• work in REPL?

story-cards.txt

- nice output stream (files)
- software synthesis
- symbolic regression
  - sine
  - cake
  - birthday
  - some big function
- robocode
- generational
- simple/aggregate fitnesses
- lexicase arg
- multiobjective arg
- crossover (tunable) CLI arg
- mutation (tunable) CLI arg
- fitness vs time (all) CSV
- training data
- population
- initial size
- language subset
- "getting stuck"
- if you're optimizing, stop if it's optimal
- if you don't find a perfect one…
- ALPS ("diversity presevation")
- steady state vs pure generational?

- id
- parent(s)
- genome
- score(s)
- origin

2000 gen x 1000 indiv x ~1000 instructions
10 Gb or so