Realistic Agent Movement in Dynamic Game Environments

This paper addresses the challenge of realistic agent movement in computer games, particularly pathfinding in dynamic environments where the game world can change in real time. It argues that traditional pathfinding methods such as A and navigation meshes depend on static, pre-processed representations of the map, which become inadequate as middleware physics engines enable players and agents to alter level geometry during play. The authors review real-time alternatives from robotics and game AI, including Real-time A, D, and steering algorithms such as force-based flocking and ray casting, noting the strengths and limitations of each in game contexts.

The central proposal is to use machine learning to give agents real-time awareness and adaptive steering. Agents are equipped with four ray-casting sensors spanning a 90-degree field of vision, whose data feeds an artificial neural network that controls movement. The network's weights are evolved through a genetic algorithm guided by reinforcement-style scoring. The approach was implemented in a test bed built on the Quake 2 engine, using a custom C++ AI library and graphical tools that let the user adjust evolution and network parameters during training.

Results were reported qualitatively rather than through extensive statistical evaluation. Agents learned to head toward a goal in obstacle-free space and, after the authors introduced discrete training intervals and specially designed "bot boot camp" maps, learned to combine goal-seeking with obstacle avoidance without prior map knowledge. The reported paths were not smooth, and the authors found the networks difficult to train, with a single hidden layer proving insufficient for the combined task. The paper reports that the system could sustain thousands of agents at 30 frames per second, and it concludes that neural networks could underpin a low-level real-time pathfinding API, with future work targeting smoother paths, hybrid networks, and reducing waypoint requirements.