Open-Endedness

From the AI Research Compass wiki

Open-Endedness: The goal is systems that keep generating novel, increasingly complex problems and solutions forever, not ones that converge on a fixed objective.

Beliefs

Open-Endedness looks at natural evolution and human culture and sees processes that never finish: they keep inventing new niches, new artifacts and new challenges, each built on the last. Most machine learning, by contrast, optimizes a fixed objective and stops when it converges. Open-endedness wants algorithms that co-generate their own problems and solutions, that value novelty and interestingness rather than a single score, and that produce an endless stream of stepping stones. Sometimes, its adherents point out, the best way to reach an ambitious goal is to stop aiming at it directly.

It sits just right of the middle of the Language-World axis, since open-ended systems range from evolving robot morphologies to language-model-driven idea generation, and a little below centre on the Harness side, because open-endedness is a property of an ongoing process wrapped around learners rather than of any single trained network. Its paradigm level is Beyond Revolutionary, the far Paradigm-shift end of the chart.

Arguments

  • Novelty search showed that ignoring the objective and rewarding behavioral novelty can solve deceptive problems where objective-driven search gets stuck.
  • Co-evolving environments and agents, as in POET, produced challenges and solutions that could not be reached by training on the hard environment directly.
  • Foundation models can supply a human-like notion of “interestingness”, giving open-ended systems a practical way to choose what to explore next.

Critiques

  • Open-endedness is hard to measure; without a fixed objective, it is unclear how to tell real progress from aimless drift.
  • Demonstrations so far remain bounded by their simulators, and no system has yet shown truly unbounded complexity growth.
  • Benchmark-minded rivals see it as a beautiful philosophy that rarely produces a deployable capability.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +0.04, y ≈ −0.46, paradigm ≈ +1.00.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data