Balancing multiplayer games design using deep learning techniques

Multiplayer games rely on progression systems to correctly pace the rise in difficulty alongside the player’s understanding of the game. This process should occur seamlessly, with a constant loop of the player improving as the game becomes more challenging. Well-structured progression allows the player to remain motivated and tackle more vigorous opponents sequentially. Multiplayer games rely on progression systems to ensure fair and fun games between different types of players. Variability factors include players with different amounts of playtime, mechanical, and/ strategic skills in the game. Testing the balance and consistency of progression systems within online games is challenging for game studios.

This dissertation proves that player replacements created using deep reinforcement learning techniques can successfully test the underlying design of multiplayer progression systems, including skill-based, economic, and level progression. Reinforcement learning agents can tackle harder games than existing techniques while exhibiting human-like behaviour in nontrivial social situations such as cooperation and competition.