It's unclear why you single out the greedy algorithm; there are many different algorithms for combinatorial optimization, the greedy algorithm (or rather, greedy-like algorithms, also known as myopic algorithms) being only one of them. That said, I have a positive answer and a negative answer for you.
Positive answer. Consider the problem of maximizing a non-negative submodular function under a cardinality constraint. In this problem, you are given a submodular function $f$ over a domain $U$, and a bound $k$. Your task is to find a subset $S \subset U$ of size at most $k$ such that $f(S)$ is as large as possible. This problem is hard (more on that below), and so we make do with approximation algorithms. The greedy algorithm has no guarantees in this regard, but a simple modification of it gives a (probably non-optimal) $1/e$ guarantee. See Buchbinder–Feldman–Naor–Schwartz.
Negative answer. In what sense is the problem of maximizing a non-negative submodular function under a cardinality constraint hard? One such sense is hardness in the value oracle model. Suppose that an approximation algorithm is given oracle access to the function $f$. Then if $f$ always produces a $1/2+\epsilon$ approximation, there must be some $f$ and $k$ for which it calls the oracle exponentially many times (in $n = |U|$), a result due to Vondrák. This hardness result holds even for randomized algorithms (with any constant success probability). So randomization has its limits as an ingredient.
Does randomness help? Take the closely related problem of unconstrained submodular maximization – the case $k = n$ (recall $n = |U|$) of the problem considered above. The algorithm of Buchbinder–Feldman–Naor–Schwartz yields the optimal 1/2 approximation, but is randomized. They also give a 1/3 approximation algorithm which is deterministic, matching previous guarantees of Feige–Mirrokni–Vondrák. Can deterministic algorithms achieve the randomized guarantee? (Rumour has it that the answer is: Yes.)
Can randomization help? Under the widely held belief that P=BPP, there cannot be any discrepancy between randomized and deterministic polytime algorithms. However, this holds only in the model in which the input is given explicitly. In the value oracle model, the input is given implicitly as an oracle, and so it could be that randomized algorithms outperform deterministic ones. Indeed, it is easy to come up with specific tasks which are much easier for randomized algorithms. The question is whether unconstrained submodular maximization is one of these examples.