Page 27 - Read Online

Zander et al. Complex Eng Syst 2023;3:9 I http://dx.doi.org/10.20517/ces.2023.11 Page 9 of 16

if thrust-dir is reverse and abs-vel-angle is forward and speed is fast then thrust is fast-reverse
if thrust-dir is reverse and abs-vel-angle is forward and speed is slow then thrust is reverse
if thrust-dir is reverse and abs-vel-angle is behind and speed is slow then thrust is slow-reverse
if thrust-dir is reverse and abs-vel-angle is behind and speed is fast then thrust is stop

if thrust-dir is forward and abs-vel-angle is forward and speed is fast then thrust is stop
if thrust-dir is forward and abs-vel-angle is forward and speed is slow then thrust is slow-forward
if thrust-dir is forward and abs-vel-angle is behind and speed is slow then thrust is forward
if thrust-dir is forward and abs-vel-angle is behind and speed is fast then thrust is fast-forward

Figure 5. Antecedent fuzzy sets, optimized TSK consequents, and associated rules for determining ship thrust.

Figure 6. Example ANFIS Structure [53] with 2 antecedents and 1 consequent.

than using fixed constants for the parameters of the antecedents and the consequences. Towards this end, the
ANFISiscombinedwithaNNtolearnrepresentationsfromtheinputs,whicharethenfuzzifiedbythenetwork.
Outputs of the ANFIS layers are calculated by multiplying the fuzzified output of the NN by antecedents of
rules. Then, the antecedents are multiplied by learnable parameters and summed to form the consequences.
In the case of multiple-consequent models where the output dimension is more than a single value, we used
a modified ANFIS structure: the Multioutput Adaptive Neuro-Fuzzy Inference System (MANFIS) [52] . This
method creates a separate rule base for each output dimension but follows the defuzzification process as a
typical ANFIS. This process is shown in Figure 8.

A NN of abstract shape is used and can be configured as desired. After passing the input through the NN, the