PUBLISHED PAPERS #02.01
| Rufat Mammadzada, Rahim Mammadzada. The Impact of Task Complexity on Agent Training Efficiency in Reinforcement Learning Scenarios |
|---|
| Abstract. This study investigates the effects of increasing task complexity on the training and performance of intelligent agents in various reinforcement learning scenarios. Beginning with a basic navigation task where the agent has precise knowledge of object coordinates, progressively challenging scenarios that test the agent's adaptability and decision-making abilities are introduced. Heuristics and sensory mechanisms, such as ray-casting and real-time feedback, are utilized along with the division of the task into smaller episodes to evaluate how agents optimize their performance under different conditions. The scenarios range from enhanced navigation with additional movement axes to sensor-based navigation without direct coordinate knowledge, and finally to dynamic environments with multiple rewards, time constraints, and competitive interactions between agents acting as hunters and prey. Each scenario is designed to incrementally increase complexity, requiring the agent to employ more sophisticated strategies and adapt Its behavior in response to new challenges. The findings provide insights into the scalability of reinforcement learning techniques and the ability of agents to handle increasingly complex tasks, offering valuable implications for agent learning in complex scenarios. |
| Keywords: Reinforcement Learning, Unity3D, Scenario, Agent, Reward |
 Download PDF |
| DOI: https://doi.org/10.30546/MaCoSEP2025.034 |