Abstract
Aim: Navigating dynamic and adversarial environments remains a key
challenge in artificial intelligence, with applications in robotics and autonomous
systems. This study explores strategic decision-making, planning, and
adaptability in the Maze Craze game environment, inspired by classic Atari
games. The objective is to train agents to efficiently navigate the maze,
capture adversarial entities, and use environmental modifications to optimize
performance. Method: A novel reinforcement learning framework was
developed using central Q-learning (super brain agent) for cooperative
multi-agent reinforcement learning (MARL). The framework trains agents
within the PettingZoo library, which provides a structured environment for
developing AI agents in Atari 2600 games. Agents strategically manipulate
their surroundings, balance short-term and long-term goals, and interact with
adversarial entities to enhance learning. Result: The study demonstrates that
the super brain agent, utilizing central Q-learning, enables cooperative agents
to efficiently navigate adversarial environments. The agents successfully learn
to balance exploration and exploitation, adapt strategies dynamically, and
utilize environmental modifications like passable walls to confuse opponents
and optimize their paths. This approach improves the overall performance of
agents in complex, partially observable mazes. Conclusion: The findings
contribute to reinforcement learning research by advancing understanding of
agent behaviour in adversarial settings. The study highlights the potential of
multi-agent cooperation, strategic environmental manipulation, and advanced
decision-making frameworks. The results pave the way for developing more
robust AI systems capable of thriving in unpredictable environments, with
applications in robotics, autonomous navigation, and intelligent game AI.

Key words: adversarial environments, Central Q-learning, Super brain agent, Multi-agent reinforcement learning, Atari environments, Reinforcement Learning, Episodes, States. Actions