Knowledge of agents should be private, i.e. each agent should have a different set of knowledge items. For example, real people only know a relatively small subset of the street network (``mental map''), and they have different knowledge and perception of congestion (e.g. [14]).
This opens the door for the use of Complex Adaptive Systems methods (e.g. [15]). Each agent has a set of strategies from which to choose, and indicators of past performance for these strategies. The agent normally choses a well-performing strategy. From time to time, the agent choses one of the other strategies, to check if its performance is still bad, or replaces a bad strategy by a new one.
This approach divides the problem into two parts (see also [16]):
The reason for this behavior is that the router miscalculates at which time it expects travelers to be at certain locations - specifically, it expects travelers to be much earlier at the location shown in the plot. In consequence, the router ``thinks'' that the freeway is heavily congested and thus suggests the side road as an alternative. Without an agent data base, the method forces the travelers to use this route; with an agent data base, agents will discover that it is faster to use the freeway.
This means that the true challenge is not to generate exactly the correct routes, but to generate a set of routes which is a superset of the correct ones [16]. Bad routes will be weeded out via the performance evaluation method. For more details see [19]. Other implementations of partial aspects are [20,21,22,14].
The way we have explained it, one will probably assume that each
individual has computational memory to store his/her plan or plans.
The memory requirements for this are of the order of
, where
is the number of people in the simulation, is
the number of trips a person takes per day, is the average
number of links between starting point and destination, and
is the number of options remembered per agent. For
example, for our Switzerland simulations with a network of
links, we have
,
,
, and
, which results
in
Since this is a large storage requirement, many approaches do not store plans in this way. They store instead the shortest path for each origin-destination combination. This becomes affordable since one can organize this information in trees anchored at each possible destination. Each node in the network has ``signposts'' for which way to go for any possible destination; a plan is thus given by knowing the destination and following the ``signs'' at each intersection. The memory requirements for this are of the order of , where is the number of nodes of our network, and is the number of possible destinations. is again the number of options, but note that these are different options per destination, so different agents traveling to the same destination cannot have more than different options between them.
Traditionally, transportation simulations use of
the order of 1000 destination zones, and networks with of the
order of 10000 nodes, which results in a memory requirement of
The problem with this second approach is that it explodes with
more realistic representations. For example, for our simulations
we usually replace the traditional destinations zones by the
links, i.e. each of the 28622 links is a possible destination.
In addition, we need the information time-dependent. If we assume
that we have 15-min time slices, this results in a little less
than 100 time slices for a full day.
The memory requirements for the network-oriented plans storage now
become
|