Offloading AI for peer-to-peer games with dead reckoning

In this paper, we study the performance of an offloaded AI agent with increasing network latencies and demonstrate that dead reckoning is effective in mitigating the observed degradation. Dead reckoning refers to a class of algorithms typically employed to predict the state of objects in existing games to mitigate the effects of game lag and improve player experience. For a deployed realtime tank game, we found that increasing latencies will cause gradual degradation to the performance of an AI agent and the performance is severely degraded when latencies reach about 300 ms. We show that a simple implementation of dead reckoning is able to delay the onset of performance degradation for round-trip latencies up to 250 ms and is relatively robust to network jitter and packet loss. Since the observed average latency within the continental North America is approximately 55 ms and inter-continental latencies are in the vicinity of 250 ms, our results demonstrate that it is feasible to offload AI to client machines. Most importantly, our method is practical because it does not require much additional code and it allows offloaded AI agents to be developed in a network-oblivious manner similar to what is presently done for server-based AI.