Page 24 - Read Online

Page 6 of 16 Zander et al. Complex Eng Syst 2023;3:9 I http://dx.doi.org/10.20517/ces.2023.11

Figure 1. Example of the Asteroid Smasher environment.

Figure 2. Simplified overview of the fuzzy inference system’s threat inputs and behavioral output.

While the primary metric for performance concerned the total number of asteroids destroyed, other metrics
included accuracy, number of lives lost, and execution speed. More importantly, the explainability of the
system proved crucial as the motivating factor for the test case. In pursuit of balancing AI explainability with
performance, TSKsystemslendingthemselvestovisualizationanddescriptionsthroughnaturallanguagewere
developed and optimized via the described learning algorithm.

3.2.2. System description
We created a base model comprised of fuzzy systems to serve as a foundation for optimization.

To balance score and execution time, each ship tracks distances and angles to immediate threats and potential
targets. These values take into account considerations such as screen wrap and object size. More importantly,
theyserveasinputstofuzzysystemsthatdetermineshooting, turning, andthrustbehavior. Ageneraloverview
of this process is shown in Figure 2.

3.2.3. Learning algorithm
The employed algorithm for optimization approximates gradient descent by iteratively running scenarios and
altering TSK outputs to more closely resemble the best performer. This requires establishing a performant
objective function and additional hyperparameter tuning; the range of possible TSK output values, number
of iterations, and learning rate all vary based on configuration. However, the theoretical upside is a more
systematic approach to optimal performance when compared to approaches such as genetic algorithms. The
approach used for this project is described in Algorithm 1.