Information Technology Reference
In-Depth Information
The main difference between the sequential and parallel simulation principle is about
the tick time. In MAS philosophy, one considers that time changes along environments
changes. In the sequential mechanism, the tick time changes after each order sent by
an agent. This means that an agent is always able to see other agents' decisions before
making its own decision. In the parallel mechanism, the tick time changes after one
decision round processed over all agents. It is a way to simulate parallelism between
the agents. This is the principle followed by the Conway's game of Life [16].
However choosing one of these arrangements is just a modelling choice. Any of
them can be obtained in a single stream architecture as well as in a multiple stream
architecture.
In a single stream architecture, one needs a specific software engineering pattern to
code the parallelism. The easiest way is to let the Market collect all the orders before
their execution. It is a way to simulate simultaneity in agents decisions. If one needs to
ensure fairness among agents, ATOM uses a loop to give the talk to all agents -
”equi-
table round table”
. An agent is allowed to act only once in each talk round. Of course,
if one wants to depart from this fairness, it is sufficient to pick randomly an agent and
to offer him the possibility to decide.
Note that an agent can decline the possibility to speak, hence, even if we have a
single stream, we can easily simulate different talking frequencies. A possibility to ex-
press an intention does not necessarily imply that a new order is issued. Since agents
are autonomous, they can evaluate their positions every round, modify trading rules ac-
cording to new market conditions. Developing an agent that sends twice less orders than
the others can be made by programming her behaviour such as she will decline word
on odd turns (keep unchanged position), while others accept to talk each time they have
the possibility to do so.
The main advantage of the single stream architecture is that the designer can repro-
duce perfectly all the experiments. He keeps the control on agent's talk. We consider
that it is the best way to build and test experiments.
In a multiple stream architecture, parallelism is obvious, but the designer does not
have the control about the talking order of agents. This order is defined by the Operating
System, and of course, it can produce biases in simulations. Nevertheless, one partic-
ularity of this approach is that the time is given in seconds - real time. It is also easy
to express the different trading frequencies for different agents, similarly with what is
described above. If there is no synchronization mechanism between the streams, the
simulation is unfair, an agent can talk twice more than another one. In a fair simulation,
one just has to put a synchronization pattern like a Cyclic Barrier to grant this property.
This architecture is preferable if one wants to include humans in the loop.
As ATOM is a multi-asset artificial market platform, we have implemented a
”one
order for one book”
rule: during a talk round, agents are just allowed to send at most
one single order to a given order topic (i.e. one order at most per stock) within the
same ”round table discussion”. This principle helps to keep the equity in agents ac-
tions. However, notice that agents have the possibility to send several orders within the
same ”round table discussion” to several order topics: this ability is simply constrained
by the
”one order for one book”
rule. If the ASM is settled such as it runs a multi-stock
