A Definition of Open-Ended Learning Problems for Goal-Conditioned Agents — part 4/4

Cleaned-text code points [47245, 48382). Bibliography omitted here; source.md includes it.

As a consequence of our choice, the main limitation of our work is that our definition of OEL problems does not imply any form of performance progress from the side of the agent. An intuition is that the agent should maximize some performance measure on the behavioral goals it has already seen so far. Thus, to go further, we should characterize a goal discovery process, explain how an agent may discover new goal spaces Pong et al., (2019) , maximize its competence in these goal spaces Santucci et al., (2016) , introduce representational redescription processes Doncieux et al., (2018) , abstraction capabilities Konidaris, (2019) ; Shanahan and Mitchell, (2022) , and even creativity Boden, (1998) . Finding adequate performance measures for all these further capabilities is left for future work, and trying to integrate all these capabilities into a common framework may reveal some deeper limitations of the definitions we have proposed here.

## Acknowledgements

This work has received funding from the European Commission’s Horizon Europe Framework Program under grant agreement N o N^{o} 101070381 (PILLAR-robots project).