MAT109 · Erich Prisner · Franklin College · 2007-2011

Doctor Location Games

Prerequisites: Chapters 1, and 2.

In this chapter and its sibling, we discuss the simultaneous versions of different location games played on undirected graphs.

Note to the Teacher: ............
Excel File

Like a digraph, an undirected graph consists of vertices. But instead of being connected by arcs with directions, vertices are connected by undirected edges that can be used both ways. Usually they are visualized by curves connecting the two vertices. Vertices connected by an edge are called adjacent. Note that edges may cross, like in the graph to the right. There is an edge connecting vertices 4 and 5, and another one connecting vertices 3 and 6, but none connecting vertices 4 and 6. Therefore vertices 4 and 6 are not adjacent, nor are vertices 3 and 5.

We use graphs for modeling transportation networks. The vertices represent the villages, and the edges the roads between the villages. Two villages are called adjacent if they are connected by a direct road. Crossings between edges should always be thought as one bridging over the other, therefore it is not possible to change the roads there.

1. Doctor Location

Doctor Location Game: On a small island, the different villages are connected by a system of streets. Two medical doctors, Ann and Beth, are about to settle. On average, every one on the island sees a doctor once a year, and always chooses the nearest one. The measure for distance is the number of roads one has to travel to go there. In case of a tie, half of the citizens of that village go to Dr. Ann and the other half to Dr. Beth.
In the simultaneous version of the game, both Ann and Beth are forced by the island government to submit a proposal for their location simultaneously. What should they choose?

Here you can play the game with different graphs: Graph 2, Graph 4, Graph 5, Graph 6, Graph 8, Graph 11, Graph 15,

Graph 1: Only one Nash: 5 versus 5.
Graph 2: Only one Nash: 2 versus 2.
Graph 4: Everyone of 1,2,3,4 versus everyone of 1,2,3,4.
Graph 5: Everyone of 2,3,5 versus everyone of 2,3,5.
Graph 8: too large
Graph 9: One Nash: 4 versus 4
Graph 10: One Nash: 9 versus 9

Note that the game is a constant-sum game, since each islander goes to one doctor once a year. Thus the sum of the patients of both doctors equals the number of islanders. Note also that constant-sum games behave like zero-sum games---we just have to subtract half of the constant sum from each payoff to get a zero-sum game where both players would still play the same. As has been shown in the Simultaneous Games chapter, these features of the games, 2 players, symmetry, and constant-sum, imply that the payoffs in each Nash equilibrium are equal (half the number of vertices) for each player, and that every maximin move of Ann achieving a guarantee of half the number of vertices versus any such a maximin move for Beth forms a Nash equilibrium.

1.1 An Example Graph

Take as an example the graph displayed to the right, Graph 6, and let me explain how the payoff matrix of the game is derived.

Continuing in this way, and keeping in mind that the game is symmetric, one can get the whole bimatrix of the game, which looks as follows:
 1  2  3  4  5  6  7 
 1  7/2, 7/2  7/2, 7/2   4, 3   9/2, 5/2  7/2, 7/2   4, 3   9/2, 5/2 
 2  7/2, 7/2  7/2, 7/2   4, 3    4, 3   7/2, 7/2   4, 3   9/2, 5/2 
 3   3, 4    3, 4   7/2, 7/2   4, 3   7/2, 7/2   4, 3    4, 3 
 4  5/2, 9/2   3, 4    3, 4   7/2, 7/2  5/2, 9/2  7/2, 7/2  7/2, 7/2 
 5  7/2, 7/2  7/2, 7/2  7/2, 7/2  9/2, 5/2  7/2, 7/2  9/2, 5/2   4, 3 
 6   3, 4    3, 4    3, 4   7/2, 7/2  5/2, 9/2  7/2, 7/2  7/2, 7/2 
 7  5/2, 9/2  5/2, 9/2   3, 4   7/2, 7/2   3, 4   7/2, 7/2  7/2, 7/2 

The Maximin moves for both players are moves 1, 2, and 5, and the corresponding values, the security levels, are 7/2 for each. Actually the four other moves 3, 4, 6, and 7 are weakly dominated by each of these moves. There are nine pure Nash equilibria, the shaded cells. As noted above, all of them must have payoffs of 7/2 for both players (but not all such fair outcomes, where both get 7/2, are pure Nash equilibria). Note also that every combination of choices of villages 1, 2, or 5 results in a Nash equilibrium.

1.2 No (pure) Nash equilibrium?

All the examples above have Nash equilibria, but this isn't necessarily so. The graph shown below has no (pure) Nash equilibrium.

1.3 How good are these Nash solutions for the public?

The villagers of the island are not players of that game, since they don't make decisions. Still there is a payoff for them, having a doctor close to their home or not. The sum of all lengths of all paths from all vertices of the graph to the nearest doctor could be seen as a measure of how good the location of the doctors is to society---the smaller this number, the better. Let's call this number the distance sum, given the doctors locations. Finding a placement of two doctors minimizing the distance sum is tedious but not too difficult---all we have to do is to compute this distance sum for every possible combination. In the graph above, Graph 6, if both doctors go to vertex 1, the distance from vertices 1, 2, 3, ... and so on to the closest doctor (at vertex 1) are 0, 1, 1, 2, 1, 2, and 2. Therefore the distance sum for this placement is 0+1+1+2+1+2+2 = 9. If one doctor is located at vertex 1 and the other at vertex 2, the distance sum equals 0+0+1+2+1+2+1 = 7. If the doctors are located at vertices 2 and 5, the distance sum equals 1+0+1+1+0+1+1=5, which is lowest possible. Obviously letting the two doctors choose their location freely does not necessarily minimize this distance sum, and therefore does not maximize society's payoff.

2. Trees

Now we know that we only need to look for vertices that guarantee half the number of vertices as payoff in order to find all Nash equilibria. Still, is it possible to identify these vertices without performing the calculations but by looking at the structure of the graph? Do we know how many vertices of this type we have? Do we know which graphs have pure Nash equilibria and which one do not?

In this section, we will find answers to these questions for special graphs, namely trees. A tree is a graph where between any pair of vertices there is exactly only path. Examples are the graphs given below.

Why trees? The reason is that the maximin move vertices---and remember that these are all we have to find--- can also be described by means of another concept, the so-called branch-weight, as will be shown in the following:

Graph 13 Applet

Graph 14 Applet

  1. We need the following definitions: If we remove a vertex x from a tree, the tree falls into parts. These parts are called the "branches" at vertex x, or "x-branches", and the maximum occurring number of vertices in such an x-branch is called the branch-weight bw(x) of x. In our examples, the branch-weights of the vertices are shown and .
  2. If Ann is located at vertex x and Beth at vertex y, then all but one x-branches do not contain y. The vertices in those x-branches are customers of x. In the same way, all but one y-branches do not contain x, and the vertices in those y-branches are customers of y. Only the vertices between x and y are split between x and y. See and for the trees considered above. Now, if x and y are not adjacent, then y could improve the number of customers by moving closer towards x. The vertices in the former y-branches not containing x still belong to y, but there are some gains in the former "between" part just about in the middle between x and y. Consequently, the best response location to a location x is always either the very same vertex or one of its neighbors. Therefore from now on we consider cases of placements on the same or adjacent vertices.
  3. The case where both players choose the same vertex is not too interesting, since the payoffs are split equally in that case, see for our example. If Ann and Beth place on adjacent vertices x and y, then removing the edge between them, the tree would fall into two parts. One of them is the y-branch attached to y at x, and the other the x-branch attached to x at y. This is also precisely the partition of the vertices into those visiting Ann's office versus those visiting Beth's office. See and and for the two trees above.
  4. Now assume that Ann places her office at vertex x in a tree with n vertices. Beth can always achieve a payoff of n/2 by placing at x too, but whenever Beth places at a neighbor of x, she gets just the corresponding branch as customers. So Beth can achieve the branch weight bw(x) of x as payoff by placing on a neighbor of x, but not more. Since we have a zero-sum game, Ann's guaranteed payoff when placing at x equals either n/2 or n-bw(x), whichever is smaller. That implies that, provided there are some vertices of branch weight less than or equal to n/2, then all these vertices are the maximin moves for the players.
Next we will show that such vertices of branch weight less than or equal to n/2 exist, and that there are only few of them [Zelinka 1968]:
  1. First note that if an x-branch of weight k is attached to x at vertex y, then bw(y) ≥ n-k, since the y-branch attached to y at x contains all vertices outside of this latter branch.
  2. Therefore all neighbors of a vertex x of branch-weight less than n/2 have branch-weight greater than n/2. Moreover, since a vertex of branch-weight n/2 can only contain one branch of this size, such a vertex has exactly one neighbor of branch-weight n/2, and all other neighbors have larger branch-weight.
  3. If a vertex x has branch-weight greater than n/2, then the neighbor y of x in the large branch has smaller branch-weight, but all other neighbors z of x have larger branch-weight than x. For, the y-branch attached at x consists of x and all small x-branches, all vertices outside the large x-branch, and has therefore less than n/2 vertices. And all other y-branches are strictly smaller than the large x-branch. For the second statement, note that the z-branch attached at x contains the large x-branch of x. Therefore, starting at any vertex, one can always move to neighbors with smaller branch-weight until one arrives at some vertex with branch-weight ≤ n/2.

Therefore we have:

Theorem: In a tree, the Maximin move vertices are the vertices minimizing the branch-weight. These are either one vertex of branch-weight less than n/2, or two adjacent vertices of branch-weight equal to n/2. The Nash equilibria are all pairs of these vertices.

Therefore in the first tree example, the only Maximin move is to play vertex 6. In the second tree example, there are two such Maximin moves, vertices 6 and 8, both with a branch-weight of 6.

If we would only have one doctor to place, the good locations minimizing the distance sum are called the "median" and were first investigated by Camille Jordan for trees [Jordan 1869]. See and for the distance sums for the different vertices for our two trees. Thus the medians are vertex 6 in the first tree, and both vertices 6 and 8 for the second one. Note that these medians, the vertices minimizing the distance sum are exactly the vertices for the Maximin moves, the vertices with minimum branch-weight, in these two examples. In fact, Jordan found that the median for trees is always either just one vertex, or two adjacent vertices, and Bohdan Zelinka showed 199 years later that the median of a tree consists of exactly those (one or two) vertices minimizing the branch-weight [Zelinka 1968].

Therefore, the Nash equilibria locations for trees would be perfect for society if there would be only one doctor, but for two doctors, they are almost always suboptimal. Both doctors behave as if they would be the only doctor in the world. This is one of the cases where competition would not lead to a solution best for society. In the first example tree, for instance, both doctors will go to vertex 6. The distance sum for this equals 19. However, if one would go to vertex 2 and the other to vertex 6, the distance sum for these locations would be 9! Surprisingly, in the second tree the Nash equilibrium of vertex 6 against vertex 8 is also optimal for society, since the distance sum equals 15, the same as for locations at vertices 6 and 8, or locations at vertices 6 and 10, or at vertices 4 and 8, or at vertices 4 and 10. But this is mainly due to the tree being so small. For larger trees, the Nash equilibrium location is usually not optimal for society, not minimizing the distance sum.

3. More than one office (optional)

We change the game slightly. Each doctor maintains two offices each, in two villages, one for before noon, and one for afternoons. We assume that people can wait half a day, so all four offices are options. In case of a tie, there are very slight preferences of the population towards the doctors, split equally, so each doctor could expect "half a patient". The choices of the offices are still revealed simultaneously.

Here you can play the game with different graphs: Graph 4, Graph 6, Graph 8, Graph 11, Graph 14. An example on a selection and the corresponding payoffs for the example Graph 6 considered before is given to the right.

We still have a symmetric simultaneous 2-player zero-sum game. But the number of options for each player is now much higher than before. If the village graph has n vertices, then each doctor has n options with both her offices placed in the same village. These options are obviously dominated. Then there are n·(n-1)/2 options with both offices in different villages. Together, each doctor has n+n·(n-1)/2 options. For our 7 vertex example Graph 6, this adds to 28 options for each doctor. In the 10-vertex Graph 4, we have 55 options for each doctor, so we are looking at a 55 × 55 matrix.

But we also still have a symmetric 2-player zero-sum game. So we still know that every pair of Maximin moves of the two players yields a pure Nash equilibrium, provided these Maximin moves carry a guaranteed payoff of half the number of vertices. But if the Maximin moves have a lower guaranteed payoff, there is no pure Nash equilibrium.

4. Geometric Version (optional)

Another class of graphs that can be analyzed easily are grids like Graph 11 or Graph 15. The Maximin moves are always the vertices in the center of the grid, which is either just one vertex, or two adjacent vertices, or a small square, depending whether the dimensions of the grid are even or odd. If we look at larger and larger grids, we may arrive at a continuous game like the one here. The two players can place their office wverywhere inside a square. Note that distance here is the usual euclidean distance, different to the grid case, where the so-called "Manhattan distance" applies.

Games like this do not quite fit into our framework, since every player has infinitely many moves. Maximin moves and Nash equilibria can no longer be found by simply checking a finite list. Moreover, rather funny things can happen, like no Maximin move, for instance. Usually tools from Calculus are applied to games like this.

However, in our example, it is fairly simple to find the only existing pure Nash equilibrium. Both place their office directly in the middle of the city, since even if Beth knows that Ann is placing her office this way, she can not really exploit this information---her best response would be to place her office also directly in the middle of the city.

Variants with two or more locations for each doctor are also possible: Play the game here. An analysis of the game with more than one office may be very difficult.

More links

Exercises

  1. Find all best responses for the graph in Section 1.2. Show that there is no pure Nash equilibrium.
    ...........
  2. Vertices that are adjacent to just one vertex, that have just one neighbor, are called end vertices. Draw a tree with 14 vertices, with exactly 5 end vertices. Calculate the branch-weights of all vertices. Find the median and all pure Nash equilibria of the corresponding doctor location game.
    ...
  3. version (n,k): Do the same as in the previous exercise for a tree with n vertices and k end vertices.
    ...
  4. Project:Doctor location on MOPS: A MOP or maximal outerplanar graph is obtained from a cycle by adding diagonal edges that do not cross each other until no more edges are possible to draw without crossing. An example is given to the right. Can you describe how many Nash equilibria are possible for doctor location games on mops? Try to start by analyzing a number of small and a little larger examples.

  5. Project:Assume a graph consists of a cycle, and three more vertices, each one adjacent to just one other vertex on the cycle. What can be said about the Nash equilibria in doctor location games on such graphs?

    ...

Projects

  1. ......

    ...
  2. .....
    ...
  3. .....
    ...
  4. .....
    ...
  5. .....
    ...