Implementing a Doctor Agent with Reinforcement Learning in Unity

Deep reinforcement learning has emerged as a powerful approach to solve complex decision-making problems in dynamic environments. In this study, an intelligent agent was developed to act as a virtual doctor in a simulated game environment, tasked with finding and treating injured players in order of their health priority. The agent was trained using reinforcement learning and was rewarded for correct treatments and penalized for errors or obstacles. The initial training method was based on the proximal policy optimization (PPO) algorithm, while soft actor-critic (SAC) and POCA were also tested for comparison. To evaluate the performance, three indicators were considered: cumulative environment reward, policy value estimation, and policy extrinsic reward. Experimental results showed that PPO consistently achieved the highest performance in all metrics, while POCA outperformed SAC. These findings demonstrate the efficiency and stability of PPO in healthcare-oriented game simulations. Future work may expand the agent's capabilities by introducing new actions such as jumping and self-defense, as well as testing alternative algorithms such as DDPG, A۳C, and TD۳ to further increase the adaptability and learning efficiency.

کلیدواژه ها:

Reinforcement Learning ، Unity ML-Agents ، Non-Player Characters ، Autonomous Agents ، ۳D Environments