Monsters:
k
CPUs:
l
Radius:
m
T
ijklm
= µ +
i
+
j
+
k
+
l
+
m
+
()
ij
+ ()
ik
+ ()
il
+ ( )
im
()
jk
+ ()
jl
+ ( )
jm
+ ()
kl
+
( )
lm
+ ( )
lm
+
ijklm
The error
ijklm
come from stochastic elements of the experiment (e.g. cpu
load) in addition to higher order interaction effects.
Calculations was performed using Minitab's General Linear Model (output
in section 6.2). By selecting a confidence level of 5%
1
, the following factors have
significant effect (checking the P-value):
1. MarkovKillers
2. PlanAgents
3. Monsters
4. CPUs
5. Radius
6. MarkovKillers*Radius
7. PlanAgents*Radius
8. Monsters*Radius
The main factors were likely to have effect, but that 2-factor interactions
involving the vision radius are not directly obvious.
The significant main and interaction factors together explain approximately
97.6% of the model based on calculations with Sum of Squares, the exact ex-
pression is (1.0 - 206214/211197)*100. The rest of approximately 2.4% is due to
true variance from stochastic elements in the experiments.
A plausible explanation of the high main effect of vision radius (fig 1) is
that the amount of vision processing increases proportional to the square of
the vision radius (processing area = · (visionradius)
2
). Since adding more
agents and increasing the vision radius rapidly increases the load of an agents
vision algorithm (i.e. sees a larger area with higher density of other agents), the
significant 2-factor interactions involving agents and vision radius (fig 1) makes
sense.
The main effects satisfy the initial assumptions, but the 2-factor interactions
involving vision radius were quite surprising at first.
There is no need to perform more experiments in order to deal with alias
structures since the experiment is a based on full factorial design (i.e. disabling
alias structures).
1
not the same as in the model above!
Paper D
95