Page 35 - Read Online

Page 16 of 22 Ernest et al. Complex Eng Syst 2023;3:4 I http://dx.doi.org/10.20517/ces.2022.54

Figure 13. The three frames from the 2-adherent marine firing control scenario are (Left) two marines with six wounded zerglings ap-
proaching, (Center) halfway through scenario with three remaining zerglings, and (Right) the scenario complete with both marines alive
due to intelligent targeting.

2 =□(( ℎ < 0.15) ∧ ( > 0.0)
(5)
→ ( < 0.25))

Despite not being explicitly directed to, the resultant chromosome received from training was proven through
JKIND to already adhere to this specification, with the resultant output confirming correctness across all input
values. However, a relevant scenario was still created and is shown in Figure 13. Within this mission there are
only two marines on the friendly forces and six zerglings. However these marines are at full health and the
zerglingsareeachwoundedtothepointwheretwoshotsfromthemarineswoulddefeatthem. Thus, intelligent
target assignment is required to win, but victory is possible.

While the training process did generate a model that was adherent to this specification, the confirmation of
specification adherence across all possible scenarios is still worthwhile.

3.2.3. Safety specifications 3 and 4
The last two specifications investigated within this study are related, both dealing with the fire control of the
siege tank. Specification 3 again being ”If a friendly unit would take splash damage from a tank shot and the
lowest health friendly in splash range is between very low and low health, only fire at the target if it would do
very high damage to the enemy” and able to represented as in Equation 6 and in LTL in Equation 7.

3 = ( ℎ > 0.0) (0.40 > > 0.15) ( ℎ . < 0.75)
(6)
ℎ ( < 0.50)

3 =□(( ℎ > 0.0) ∧ ( > 0.15) ∧ ( < 0.40)
(7)
∧ ( ℎ < 0.75) → ( < 0.50))

This is due to the fact that the siege tank will only fire at a target if the fire bid is above 50%, so if the shot
effectiveness is below ”very high” then we do not wish to fire. This specification did fail with the post-training
chromosome, and produced a counterexample as can be seen in Table 3. The counterexample displayed here
demonstrates that our AI model was perhaps close to adhering to the specification; the shot priority value
was not far above the maximum threshold of 0.50. For this specification there are three FISs to consider, but
given the context of the specification a general reduction in the output MFs of the Shot Safety FIS was selected
to modify this FIS until it was adherent. However, by changing the MFs and RBs in any of the three FISs