BioelectricaAlterspan Labs
Minds of small things  ·  Article

Intrinsically motivated goal exploration (IMGEP), novelty search and quality diversity

Novelty search, quality diversity and intrinsically motivated goal exploration processes (IMGEPs) are machine learning techniques built to explore a behaviour space or goal space efficiently.3 They matter to this field because the same tools can map what a self-organising system, such as a gene regulatory network, is able to do.56

Earliest held
2019, Levin M
Most discussed in
AI-driven automated discovery tools reveal…, 2025
In the library
18 passages in 3 works
Rewritten
2026-11-01
18501875190019251950197520002025Levin M, 20191Levin M, 20192Etcheverry M, Moulin-Frier C, Oudeyer…, 20253Etcheverry M, Moulin-Frier C, Oudeyer…, 20254Etcheverry M, Moulin-Frier C, Oudeyer…, 20255Etcheverry M, Moulin-Frier C, Oudeyer…, 20256Etcheverry M, Moulin-Frier C, Oudeyer…, 20257Etcheverry M, Moulin-Frier C, Oudeyer…, 20258
01

Goal space as a boundary

Levin (2019) used the term goal space in a different setting, the cognitive reach of an agent. He wrote that the edges of an agent's goal space form a sort of computational light cone, beyond which its cognitive system cannot operate.1 As examples he gave the tick, with a small cognitive boundary, little memory or prediction in time, and very local sensing and acting.2 He did not describe a search algorithm there.

02

Diversity-driven methods

Etcheverry, Moulin-Frier, Oudeyer and Levin (2025) name three techniques: Novelty Search, Quality Diversity and IMGEP. They cite Lehman and Stanley for novelty search and Cully et al. and Pugh et al. for quality diversity.3 They describe all three as designed to explore a behaviour space or goal space, which is a predefined or learned model of the overall state space. According to the same authors, IMGEPs were first developed in robotics for learning inverse models of highly redundant mappings, citing Baranes and Oudeyer (2013).4

03

IMGEP applied to networks

Etcheverry and colleagues report that IMGEPs have helped discovery in complex self-organising systems, citing Grizou, Falk, Reinke and earlier work by Etcheverry.5 In their 2025 paper they use an IMGEP to set the initial states of a gene regulatory network. The aim is to maximise the diversity of discovered endpoints within a fixed budget of experiments.6 A goal generator samples goals using intrinsic motivation, such as novelty or learning progress.7 A goal-conditioned policy then proposes intervention parameters to reach the current goal.8

SourcesEach quotation was checked word for word against the passage it opens.
  1. The edges of a given Agent’s goal space define a sort of “computational light cone” – the boundaries beyond which its cognitive system cannot operate.Levin M, 2019 · The Computational Boundary of a "Self": Developmental Bioelectricity Drives… · open at passage 34
  2. a tick has a relatively small cognitive boundary, having very little memory or predictive power in the temporal direction, and sensing/acting very locallyLevin M, 2019 · The Computational Boundary of a "Self": Developmental Bioelectricity Drives… · open at passage 34
  3. recent diversity-driven machine learning techniques such as Novelty Search (Lehman and Stanley, 2008; Lehman and Stanley, 2011), Quality Diversity (Cully et al., 2015; Pugh et al., 2016)Etcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 24
  4. IMGEPs, which were originally developed for the learning of inverse models of highly-redundant mapping in robotics context (Baranes and Oudeyer, 2013)Etcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 24
  5. were recently shown to successfully assist discovery in complex self-organizing systems (Grizou et al., 2020; Falk et al., 2024; Reinke et al., 2020; Etcheverry et al., 2020)Etcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 24
  6. we propose to use an IMGEP to control GRN initial states and maximize the diversity of discovered endpoints {z∈Z} within a limited budget of N experimentsEtcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 25
  7. a goal generator module G that samples goals from the goal space based on intrinsic motivation incentives (e.g. to promote novelty or learning progress)Etcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 25
  8. a goal-conditioned optimization policy Π that generates candidate intervention parameters to achieve the current goalEtcheverry M, Moulin-Frier C, Oudeyer PY, Levin M, 2025 · AI-driven automated discovery tools reveal diverse behavioral competencies of… · open at passage 25
Linked ideas
Covers intrinsically motivated exploration of behaviour, the motivation principle that drives the goal generator in an IMGEP.
Concerns measuring diversity of behaviour, which is what these methods aim to maximise across discovered endpoints.
Gene regulatory networks are the system whose behavioural repertoire Etcheverry and colleagues explored with an IMGEP.
Novelty search and quality diversity belong to the evolutionary computation family discussed there.
Levin's goal space and the goals sampled by an IMGEP both rest on what it means for a system to pursue goals.
Exploration for its own sake, as in novelty-driven search, sits against the explore-exploit tradeoff treated in that article.
Curiosity search is the diversity-driven strategy set against optimisation-driven baselines under the same budget.
Where it is discussedPassages matching IMGEP, intrinsically-motivated goal exploration process, novelty search, quality diversity
2025AI-driven automated discovery tools reveal diverse behavioral competencies of… · Etcheverry M, Moulin-Frier C, Oudeyer…16
2019The Computational Boundary of a "Self": Developmental Bioelectricity Drives… · Levin M1
2022Biology, Buddhism, and AI: Care as the Driver of Intelligence · Doctor T, Witkowski O, Solomonova E…1