background image
2
Choice of Factors and Levels
Five factors believed to have effect on the outcome of the experiment are chosen
1. number of agents of type MarkovKillers (per CPU)
2. number of agents of type PlanAgents (per CPU)
3. number of agents of type Monsters (per CPU)
4. number of CPUs
5. vision Radius of agents (for all types)
Table 1. Factor levels
Level MarkovKillers PlanAgents Monsters CPUs Radius
1
500
500
50
10
10
2
1000
1000
200
17
20
The number of CPUs and monsters are likely to be very significant for the
runtime since they are controlling the amount of movement that MarkovKillers
and to lesser extent PlanAgents do. Movement between CPUs leads to (timewise
costly) network traffic. Since monsters can't move, they are not likely to increase
network traffic, and with computationally cheap Markov-based action-selection
they are not likely to increase the CPU load much either. The effect of sight
radius is uncertain, but it can potentially lead to more network traffic and more
CPU load.
Interaction between factors is likely to occur, in particular between CPUs
and monsters, since increasing the number of CPUs and monsters together will
give more network traffic due to more doors between subworlds at the CPUs.
Note that the number of agents are (initally) per CPU (i.e. not total number of
agents. This can potentially reduce the effect of the CPUs since the information
is somehow present in the factors describing the number of agents.
The selected levels presented in appendix C has been determined in discus-
sions with the implementors (MSc students) of the simulator.
Since this is a computer simulation the selected factors can be completely
controlled and set to the wanted levels. Examples of factors that can't be con-
trolled is the load on the computer cluster, but the batch job scheduler minimizes
the risk of high load from other applications or systems simultaneous with our
experiments.
Choice of Response Variable The selected response variable is the (wall
clock) time for running simulations of 100 cycles (i.e. simulating 1 second per
cycle) with the variations of levels shown in 1.
Other response variables of interest could be the load of CPUs of the com-
puter systems or type of actions performed by the agents. These are however
more computationally expensive to measure than the wallclock time.
Paper D
93

<< - < - > - >>