Page 25 - Read Online
P. 25

Page 6 of 22                   Ernest et al. Complex Eng Syst 2023;3:4  I http://dx.doi.org/10.20517/ces.2022.54






























               Figure 2. Primary scenario to be utilized for training and analysis. Consists of a stronger zergling force within close range of the human
               force. Pictured (1) siege tank, (2) medivac, (3) marines, (4) zerglings.


               A primary trainingscenario has been developed which will be utilized for reinforcement learningfor a difficult
               engagement made up of these units. A plurality of other engagements have been developed for testing and
               formal verification.


               From a raw performance perspective, the mission shown in Figure 2 is the primary performance objective to
               complete. This mission has 16 zerglings that make up the hostile forces as well as 6 marines, 1 medivac, and 1
               siege tank on the friendly team.


               This mission is such that if the in-game controllers for both forces behave natively, the human ground forces
               lose with on average at least 7 of the zerglings still alive. This mission is feasible for a human to complete with
               some forces remaining, but is very difficult and requires advanced tactics within the game. Despite expertise
               in the game, the standard ending with manual full focus on controlling the human forces has notable losses.

               2.2.2. Tree structure creation
               A GFT has been created to control these 3 specific types of human units, both individually or as an entire
               force. The general approach for these sorts of control problems is to generate an entire action plan each time
               step of the environment. As such, the GFT utilized within this study will provide significantly higher action
               throughput than what would be maintained by a human.


               Throughsubjectmatterexpertisewithinthistypeofengagement,weunderstandthatthereareafewkeycontrol
               decisions that need to be made each time step:

                • The Medivac should make healing decisions that efficiently utilize its resources and keeps biological units
                  alive as best as possible.
                • TheMarinesmustbe intelligentinhowthey spreadtheirfireagainsttheincominggroupoftargets, focusing
                  fire to eliminate enemy units efficiently.
                • The Marines optimally will utilize the ”stutter step” strategy, both to increase their effective attack speed as
                  well as to potentially minimize the total incoming damage from the enemy melee units.
                • The Siege Tank is a powerful unit and should attempt to get as many effective shots off, while minimizing
                  the harm it does to friendly units.
   20   21   22   23   24   25   26   27   28   29   30