August 17, 1999, modified March 25, 2001 & November 14, 2001
Definition #1: A voluntary action is an action by an individual whose outcome can be modified by the individual (e.g. upon request), and as such it is predicted at least as well on average by the individual in question as by anybody else.
or
Definition #2: A voluntary action is an action by an individual whose outcome can be modified by the individual (e.g. upon request), and as such it is predicted better on average by the individual in question than by anybody else.
I would say that free will is the ability of the individual to perform voluntary actions.
I say on average because in particular situations others may be able to predict behavior better. For example, if I know I'm about to hit your knee, I might know better than you that you are about to have a knee-jerk reaction. Although this case could excluded by limiting behavior to behavior enacted by voluntary muscles, this would still not preclude a stealth bomber from predicting that you will run better than you would. On average, though, a person is (at present) the best short-term predictor of what he/she will do.
Note that definition #2 suggests that if our technology ever becomes good enough to predict someone's behavior better than the person can himself/herself, we will stop saying that person has free will. I find this to correlate with my intuitive definition of free will very well: if someone can predict what I will do better than I can, it means not only that I am predictable to him/her, but more importantly, that I cannot get myself to do everything I predict I will do --since you cannot beat 100% predictability. If and when that happens, I will gladly admit I lack a completely free will.
It is of interest to observe that whether one takes the definition to be the individual's ability to predict his/her behavior to be "greater than" or "great than or equal to" anyone else's makes an important conceptual difference. In the former case, the emphasis is on unpredictability, and thus technological advances that enabled us to predict a person's behavior would eliminate mankind's free will. In the latter case, the emphasis is instead on predictability of behavior by the individual, and thus following this definition, future or present predictability of human action would have no effect on the existence of free will, insofar as it does not alter a human's ability to carry out a stated plan.
Note also that for empirical verifiability, this definition would deem a person uncapable of establishing communication with the outside world uncapable of demonstrating free will, but not uncapable of possessing it.
I am not quite satisfied with definition #1, since very simple programs that can predict their behavior perfectly would qualify as having free will. Definition #2, on the other hand, implies that the essence of free will is not the subject's ability to carry out a plan, but rather the fact that the plan is unknown to anyone else (and thus possibly the free willer's creation).