SimWB

Why simulate economic interactions?

SimWB is still under development, and there is no neat package to download. However, if you are interested to play around with it, feel free to download the binary, executable on a SUN4. The interface parts are written in Tcl/Tk so you will have to get that file too. (To download the binary you will have to choose "save link as..." in Netscape.)

simwb.tar.gz (sun binary file and tcl file)

Put the files in the same catalog and run sim3.

The program starts immediately. The world is a 100x100 surface with 5x5 rooms. To the left you can modify these parameters and press restart.

There are three different actors: buyers (red), nice sellers (yellow) and malicious sellers (black). A buyer wants at each step to buy the goods produced by the sellers. The buyer chooses a maximum price it is willing to pay and tries to find a seller charging that price (or less).

A seller is nice if his price is less than the value his goods gives to the client, yielding a net profit for the customer. Conversely, he is malicious if it is a net loss to do business with him.

To the right there are graphs showing statistics of the execution. The topmost graph shows which seller strategies are the most common. Initially there are a few of each strategy, but after a while sellers who do bad will start to disapear and other sellers will start to do better.


Initial strategies	Sellers with low price and low value become common.	The sellers die out

In the above example the agents were all using the strategy of choosing the cheapest seller in the neighbourhood. This explains why expensive sellers die out.

It is possible to change which strategy the byers use to select seller. To the right there is a panel that controls their strategy. Clicking on it changes the behaviour immediately.

Cheapest chooses the cheapest seller in the range of sight
Favorite makes the buyer remember the last seller he dealt with if the seller made him a good deal (where value > cost).
Closest simply picks the closest seller (this could be an alternative way to implement locality if the motion is restricted)
Anyone Yep, it picks anyone! (Here the sellers can pick a buyer and try to make him buy his product.)
Reputation01 or room reputation. In every room there is a list of which seller who gives the most value for a given price. When a buyer wants to buy something, he can consult this list.

Modifying the step length makes the agents move in larger steps. Often it seems to be a bad idea if this is to high and the view lenght is small (or there are many rooms), because this makes the agents have to deal with new, unknown agents all the time.

There are two more graphs to the right. The middle graph shows the number of buyers. Depending on the topology if the room this will differ. In a world with long range of sight and few rooms, there will be a smaller number of sellers (and larger number of buyers). By changing to a topology with many rooms or short sight there will be more sellers.

The lower graph shows the percentage of the transaction value made by the malicious agents. If this is zero the agents arrive at choosing good sellers, even though there can be malicious agents left.

Try running a simulation with the favorite connection algorithm. You will find that the system stabilises at a population of sellers with strategies alog a diagonal line that cuts the y axis a little above zero. These are yellow seller strategies, since they correspont to value - price > 2 (it costs 2 units to participate each time). But sellers with a low price will disapear, since they are charging too low to make enough profit to stay in the game.

Using "favorite", the overcharging sellers die out and the good ones remain.

Another experiment could be to use the anyone algorithm. Here the expensive malicious sellers will take over and the population will finally die out. (Remember, with the cheapest algorithm, it was the cheap malicious sellers who took over.)

Have fun!
Lars Rasmusson
lra@sics.se