Here's an example:
In the Mathscinet review of "Y-systems and generalized associahedra", by Sergey Fomin and Andrei Zelevinsky, you find:
Let $I$ be an $n$-element set and $A=(a_{ij})$, $i,j\in I$, an indecomposable Cartan matrix of finite type. Let $\Phi$ be the corresponding root system (of rank $n$), and $h$ the Coxeter number. Consider a family $(Y_i(t))_{i\in I,\,t\in\Bbb{Z}}$ of commuting variables satisfying the recurrence relations $$Y_i(t+1)Y_i(t-1)=\prod_{j\ne i}(Y_j(t)+1)^{-a_{ij}}.$$ Zamolodchikov's conjecture states that the family is periodic with period $2(h+2)$, i.e., $Y_i(t+2(h+2))=Y_i(t)$ for all $i$ and $t$.
That conjecture claims that an explicitly described algebraic map is periodic. The conjecture can be checked numerically by plugging in real numbers with 30 digits, and iterating the map the appropriate number of times. If you see that time after time, the numbers you get back agree with the initial values with a 29 digit accuracy, then you start to be pretty confident that the conjecture is true.
For the $E_8$ case, the proof presented in the above paper involves a massive amount of symbolic calculus computations done by computer. Is it really much better than the numerical evidence?
Conclusion: I think that we only like proofs when we learn something from them. It's not the property of "being a proof" that is attractive to mathematicians.
