Re: Transport models

[Mark H. Butler]

Hi Xueyue Huang

I'm posting this to Swarm modelling because my discussion is more suited to
that list.

> My project is about simulating travellers' behaviour in a transport
network where
> information systems are used. I designed a multiagent framework for this
project
> in which travellers, information systems, and transport networks are all
taken
> as agents. However, these agents' decision-making procedures are closely
related to each other, i.e., one agent's decision-making will be influenced
by the 
> status (for example, traffic congestion level of transport network) of
other 
> agents. Therefore, they can not make decisions without considering other
agents'status. So, my question is: in this simulation, how to make it
possible for 
> all agents to live against the same clock system such that each agent can
get 
> other agents' CURRENT status. 

Well I think there are two different issues here:

1) What structure is most suitable for your model?
2) How you can implement the model in Swarm?

I'd like to propose it is best to consider these issues independently as
otherwise you run the risk of your implementation strategy "shoe-horning"
your model i.e. you are forced to implement the model in a particular way
which may or may not reflect the system's true structure. I know very
little about transport networks as my area of study is biological systems
but I'd like to suggest some points you might like to consider in relation
to issue 1. Please excuse me if they are obvious to you. Hopefully other
people on this list will be able to elaborate issue 2.

The concept of information in complex systems is an important issue. In
particular it is important when we are dealing with agents inside the
system which have limited knowledge about the system as a whole. Formally
this is to do with the difference between exo and endo descriptions of
systems. Let's consider this in relation to transport networks:

1. In transport networks there are global information sources. An example
of this would be a radio program that broadcasts information about traffic
conditions, road closures etc. Such information sources are global in that
all agents can receive them (assuming they have radios) and they convey
information about events and conditions which may be far away from the
agent's current position. However such information sources are also prone
to errors i.e. the information may be inaccurate, subject to noise or
delays such that a traffic condition arises a considerable time before it's
existence is noted by the global information source.

2. There are also local information sources. Each agent can sense
conditions locally i.e. the speed of traffic ahead, the speed of traffic in
the opposite lane, the weather conditions (which may affect traffic
conditions) and emergency vehicles speeding past indicating accidents.

3. There are also historical information sources. Different agents tend to
follow the same routes. A particular delay one day (or over a series of
days) more cause the commuters to change routes. Agents will also know that
traffic at particular times of the day is prone to congestion so will try
to plan journeys at other times.

(For others who are interested these concepts apply to a lot of models. I'm
working on models of the liver. In that domain hormonal signals consitute
global information, intercellular signals such as calcium ions propagating
via gap junctions or prostaglandins consitute local information and
intra-cellular enzyme and protein distributions consititute a historical
information source (arguably!))

I expect that local information will be important to your model. Because
local information is important, I propose it's quite likely your model will
have to make space explict. 

Another aspect you will need to consider is constraints. I'm currently of
the opinion that most models of complex systems involve constraints, either
implicitly or explicity. It's better if our constraints are explicit as
then we fully understand the assumptions we have made in constructing our
model. Adding constraints are also a way to make our model more complex and
hence 'more realistic'. Sample constraints include:

4. The inaccuracy or delay associated with global information sources i.e.
it takes an hour before a traffic jam is reported on the radio.

5. The scope of local information i.e. can the agent judge traffic
conditions 100 metres ahead? 400 metres? More? How is this affected by the
shape of the road network?

6. The speed of vehicles in relation to weather conditions or traffic
conditions i.e. once traffic reaches a very slow speed, this has a positive
feedback affect such that even when the condition causing the delay is
removed traffic will move slowly.

7. A cost associated with receiving and interpreting information. For
example can agents get information from global or local sources and carry
out actions at the same time? Do they receive information at all times or
only at certain times?

8. The likelyhood of events not determined by the agents such as roadworks
or accidents occurring and how they are related to the global parameters of
the model.  As those of you who have played SimCity will know as road
traffic increases, road wear will increase and hence the amount of road
works.

9. The particular time of day an agent needs to make a journey.

10. The layout of roads and traffic control measures.

(Again constraints apply to other domains. In respect to the liver one
important constraint is the rapid changes in plasma glucose concentration
in relation to the delay associated with enzyme expression. This has
implications on how hepatocytes - the agents in the liver - perform their
job of controlling plasma glucose levels).

Another important aspect is heterogeneity of agents. One important
advantage of bottom-up models is that it is possible to simulate the fact
that different agents will respond to the same conditions in different
ways. Some examples include:

10. The sensitivity of the agent to global information sources. For example
how often do they listen to the radio)?

11. The sensitivity of the agent to local information sources. For example
do they modify their normal behaviour as a result of traffic conditions? Do
ignore red lights?

12. The sensitivity of an agent to learning new patterns. For example if an
agent is forced to travel to work a different way one day, do they then
follow that route from now on or do they go back to their old behaviour
once the external stimulus is removed?

13. The goals of the agent. For example different drivers may follow
different journeys. Some may follow the same journey each day, others may
follow different journeys.

(Again agent heterogeneity is important in many domains. In the case of the
liver, example differences between agents include intracellular enzyme
distributions, number of receptors and level of glycogen storage).

Here I've assumed you are just dealing with drivers as agents. It's also
possible that you have other agents in your system. For example you might
have a computerised traffic control system in charge of the traffic lights.
This makes decisions about directing traffic flow etc based on measurements
on the traffic system. The information available and the range of actions
which can be performed by such an agent are quite different to the drivers.


(This is true of other systems like the liver. For example we might want to
simulate the pancreas as well as hepatocyes because it controls glucose
metabolism via insulin and glucagon but has quite different properties and
receives different signals to hepatocytes).

The heterogeneity of the agent is also tied in with the decision making
process it uses. Let's call this decision making process the controller.
The structure on the controller is a complex issue but here are a few
important issues:

14. The controller can be symbolic i.e. we can look inside it and see how
it is reaching a decision. 

15. The controller can be subsymbolic i.e. we cannot see how it is reaching
its decision. For example we might use a neural net or a classifier system.

16. It's possible to get both types of controllers to learn. We just need
modifiable parameters in our decision process and either an optimization or
competitive strategy for updating those parameters. 

17. We can imagine a hierarchy of controllers increasing in complexity. At
the bottom we have a simple homeostatic controller which is trying to
adjust an output so it is always at a certain level like a thermostat. At
the top of the hierarchy we have a controller which can produce a plan
consisting of a series of actions which needed to carried out in a certain
order to reach a certain goal. Planning is complex as it requires the agent
to have its own internal model of the external world so it can consider
various plans, decide on the optimal plan and then carry it out. 

There must be other types of controllers in this hierarchy. Unfortunately
in the published literature on this subject there is a lot of work
considering a particular type of controller in siolation but little work
comparing different types of controllers and their properties. This would
be very helpful to your application as you need to make an intelligent
decision about the type of controller which would best suit your model.
Perhaps someone else on the list can suggest a suitable source or
reference?

Unfortunately I guess your agents will be near the top end of this
hierarchy. You need to decide on a suitable controller which allows your
agents to make journeys with some degree of intelligence and allows the
degree of heterogeneous behaviour you require.

(This is where my work gets easier. I think hepatocytes are nearer
homeostats than full blown planning, intelligent controllers. Others would
disagree with me though.) 

Okay I've written quite a lot now. I hope some of it is of use to you or to
other people on the list. If anyone has comments on this I'd be very
interested to hear them.

Mark H. Butler
address@hidden                  http://www.csc.liv.ac.uk/~mhb
L'pool Biocomputation Group   http://www.csc.liv.ac.uk/~biocomp
Postgrad/Mature Student Soc http://www.csc.liv.ac.uk/~mhb/postgrads.html







                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================