A Closer Look at Scoring

Abstract

As outlined in Section 1.4 and by Figures 1.1 and 1.4, MATSim is based on a co-evolutionary algorithm: Each individual agent learns by maintaining multiple plans, which are scored by executing them in the mobsim, selected according to the score and sometimes modiied. In somewhat more detail, the iterative process contains the following elements: mobsim The mobility simulation takes one "selected" plan per agent and executes it in a synthetic reality. This may also be called network loading. scoring The actual performance of the plan in the synthetic reality is taken to compute each executed plan's score. replanning consists of several steps: 1. If an agent has more plans than the maximum number of plans (a connguration parameter), then plans are removed according to a (conngurable) plan selector (choice set reduction, plans removal). 2. For some agents, a plan is copied, modiied and then selected for the next iteration (choice set extension, innovation). 3. All other agents choose between their plans (choice). An agent's plans in a given iteration may be considered the agent's choice set in that iteration. As a result, steps 1 and 2 of replanning modify the choice set, while step 3 implements the actual choice between options. Choice is typically based on the score; higher score plans are more likely to be selected. This is discussed in more detail in Chapters 47 and 49. For the time being, note that the three steps of replanning must cooperate for the approach to work: the plans removal step should remove "bad" plans, the innovation step should generate "good" plans, and the choice should, ingeneral,