Improving games AI performance using grouped hierarchical level of detail

Computer games are increasingly making use of large environments; however, these are often only sparsely populated with autonomous agents. This is, in part, due to the computational cost of implementing behaviour functions for large numbers of agents. In this paper we present an optimisation based on level of detail which reduces the overhead of modelling group behaviours, and facilitates the population of an expansive game world. We consider an environment which is inhabited by many distinct groups of agents. Each group itself comprises individual agents, which are organised using a hierarchical tree structure. Expanding and collapsing nodes within each tree allows the efficient dynamic abstraction of individuals, depending on their proximity to the player. Each branching level represents a different level of detail, and the system is designed to trade off computational performance against behavioural fidelity in a way which is both efficient and seamless to the player. We have developed an implementation of this technique, and used it to evaluate the associated performance benefits. Our experiments indicate a significant potential reduction in processing time, with the update for the entire AI system taking less than 1% of the time required for the same number of agents without optimisation.