The most accurate and most quantum definition is that it is the set of all maximally symmetric states (preserving the spacetime symmetries of the flat space or de Sitter space or anti de Sitter space if one allows gravity) in the extended Hilbert space – or one connected component of such a space. By the extension, I mean the formal union of all the superselection sectors.
Equivalently, one may consider the moduli space in QFT to be the set of all superselection sectors that include the maximally supersymmetric ground state.
It just happens that one may associate each such maximally symmetric state with a stationary point (minimum) of the effective action and it is indeed the 1PI effective action. We also talk about the Wilsonian, low-energy effective actions which are conceptually important but they're not an accurate tool to describe the moduli spaces; see some uncontroversial comments about the "right effective action to be used" written by my ex-adviser in his otherwise controversial paper
http://arxiv.org/abs/hep-th/0412129
It's often the case that we know exactly what decides about the fate of the moduli space. In most cases, the theory admits a classical limit and the true moduli space has to be a "deformation" of the classical moduli space – the set of solutions to the classical equations. However, the potential may be modified by quantum corrections that may lift the degeneracy and lower the dimension of the Hilbert space (or set it to zero).
Also, as the ${\mathcal N}=2$ supersymmetric gauge theories show, the quantum effects may introduce monodromy and change the topology of the moduli space etc.
This post has been migrated from (A51.SE)