Economic Perspectives I

Main Menu Previous Topic Next Topic

**Effect Of Altering Simulation Parameters**

Note: in order to run the simulation referred to in this slide, go here, to the Java Applet version. You will be directed to download the latest version of the Java plug-in.

Hopefully, by now the mechanisms behind our simulation are a bit clearer. The Agents pick among the four available Actions according to their internal probability distributions. Every other step, the Informed Agent asks for advice from the Omniscient one, and uses Bayes' rule to update her knowledge base.

As you may have noticed, Informed Agent doesn't trust that advice blindly. In fact, she knows the exact probability that Omniscient Agent recommends the true state, a number which she takes into account during her knowledge update.

You may alter the precision of Expert's (i.e. Omniscient Agent's) Advice using the slider below the performance graph. Not surprisingly, Informed Agent's performance (the slope of the green line) increases with the increasing quality of information she receives from the Expert.

Another parameter you may want to change is the Number of Updates that Informed Agent gets before the true state's identity undergoes a random change. You may think of this parameter as the inverse of the environment mutation rate. Increasing the Number of Updates tends to boost the Agent's performance, by giving her more time to learn and exploit her knowledge of the world state.

You may notice, incidentally, that increasing the Number of Updates is most effective when the Precision of Expert's advice is mid-range. Intuitively, this makes sense, especially if we consider the edge cases: when Expert's Advice is completely random, Informed Agent's knowledge doesn't improve in any case, and giving her more time to learn doesn't help. On the other hand, when the advice is flawless, a single update is all it takes to perfect the Agent's knowledge, and subsequent updates do nothing.