Actually, this topic appeared recently as a thread in comp.ai.games. The poster was asking how to do path planning with dynamic obstacles. Here you are trying to move toward the dynamic objects, rather than avoid them. The same technique can be applied.

Potential fields are a method of path planning that you can find reference to in the literature. To summarise:

Assign negative ''charges'' to each of your targets, varied by how ''attractive'' they are as Ferretman mentioned;
Assign a positive ''charge'' to all things you want to avoid, again weighted by how ''unattractive'' they are;
Assign a positive ''charge'' to your object (lets assume its the NPC sub).

The desired direction of the velocity vector of the NPC sub is the negative gradient due to the sum of all potential gradients at the point of the object.

In ''lay'' terms, think of this problem as a set of hills and valleys and the submarine is a ball. Things you want to avoid create hills with them on top. Things you want to move to create a valley with them in the middle. Place all of the hills and valleys in your domain and at the point of the sub sum the height contributions from each hill and valley (positive and negative values) to get the mean height (field potential) at that point. Now, work out the local gradient (where the sub is) and align the velocity vector with the direction that minimises the height (ie, points down the local hill). This would be the direction the ball (sub) would roll down if you let it go at that point.

Update this every time iteration of your AI code.

There are some ways to make this computationally more efficient but I wont go into them unless someone really wants to hear them.

Hope this helps,

Tim