Experiments

Left figure visualizes the busy AI and lazy AI styles learned by EMOGI, where one can find that two styles varies greatly. Specifically, during game playing, the busy AI moves a lot while the lazy AI does not.

More importantly, both styles can win the game, practically proving that EMOGI can not only learn a Game AI to win, but also control the way it wins.

The behavioral data of a Game AI during game playing is collected and calculated using I_win-rate , I_busy and I_lazy indicators (yellow, red and green color in left figure), measuring the winning rate, busy degree and lazy degree, respectively. In left figure, one can find that EMOGI not only learn extreme styles (lazy- and busy-win styles) along with two objectives, but also more nature AIs evenly distributed among them. Both busy AI and lazy AI perform extreme behavior (highest in I_busy or I_lazy), however, more nature styles are also learned by EMOGI.

Left figure (a) visualizes a combat scenario, where a easy- to-kill wizard need to beat a hard-to-kill fighter. Figure (c,d) visualize the footprint of the aggressive AI (Agg AI) and defensive AI (Def AI) learned by EMOGI, where one can find that two styles varies greatly. Specifically, during game playing, the Agg AI always attack, barely moving, producing shorter trajectories. By contrast, Def AI creates longer trajectories by running around the field to avoid the opponent.

More importantly, both styles succeed in defeating the opponent, confirming the ability of EMOGI in generating varied human-like styles.

Statistical indicators I_win-rate, I_agg, and I_def are adopted to quantitative measure the behavior of the learned styles by measuring the winning rate, aggressive degree and defensive degree, respectively. Left figure visualizes the distribution of the statistical results. Styles learned by the EMOGI are evenly distributed between two dimensions, where more varied and nature policies exists. On the other hand, both aggressive and defensive AI perform extreme behavior (highest in I_agg or I_def), however, nature styles performs more like a human.