Imagine an online coop style FPS or TDS game where say 2-6 real players fought large amounts (20-60 concurrent) of AI enemies. Could this style of game be reasonably networked over the internet using a Quake or Tribes style partial state replication model? Almost every coop game I know of uses a lockstep style networking in these cases (even games like Halo which use state replication for normal multiplayer, but lockstep for coop/firefight). My concern is that the network download (for clients) and upload (for server) overhead of each AI would end up being about the same as if a real player was in the game. Is this a real concern? Or is the limiting factor on the number of players more about how much data the players upload?

TLDR; Can large numbers of AIs be reasonably networked to a few players with a partial state replication model or is lockstep the only real solution.

Any help here would be great!

  • \$\begingroup\$ You're concerned about the bandwidth usage, correct? Are you using a peer-to-peer model or are you okay with using a server-client model? \$\endgroup\$ Commented Jun 11, 2013 at 1:44
  • \$\begingroup\$ Also, when you say "partial state replication", are you alluding to Client-side prediction? \$\endgroup\$ Commented Jun 11, 2013 at 2:03
  • 1
    \$\begingroup\$ client server model. Yes I am concerned with bandwidth. Partial state replication meaning I don't need to send the entire gamestate every frame. Only objects that need updates are transmitted using a delta system. \$\endgroup\$ Commented Jun 11, 2013 at 2:33
  • \$\begingroup\$ You may want to consider editing the title to really hone in on the question you're asking. Remember that you're not only asking questions for yourself - you're asking for future visitors too! Something along the lines of "How much do large amounts of AI enemies affect networking performance in online games?" will suffice. \$\endgroup\$ Commented Jun 11, 2013 at 3:31

1 Answer 1


Yes, online games with a large number of artificially intelligent agents is more than plausible.

When using a client-server model, the server runs a master copy of the game simulation, and the clients run their own copy if you are using client-side prediction.

Starsiege: Tribes was able was able to handle 128 human or AI players... in 1998.

The multi­player mode supports up to 128 human or AI controlled players in a single game.  Performance over the Internet drove the design of the networking model. The model supports low end modem connections and is designed to deal with low bandwidth, high latency and intermittent packet loss.


You can treat AI players just like networked players, with one large benefit - since they live on the server, they don't need to be sent any data which saves the server's bandwidth.

For a game that features many agents running around in an online open world, see Running with Rifles.

  • \$\begingroup\$ Do you know what internet speeds/bandwidths were required for 128 player matches? \$\endgroup\$ Commented Jun 11, 2013 at 3:30
  • 1
    \$\begingroup\$ I do not, but the best internet connections at the time were T1 connections. However, those were typically used by servers and people with money, like John Carmack. Most players probably had ISDN lines, which had a max speed of 128 kilobits per second. \$\endgroup\$ Commented Jun 11, 2013 at 3:43

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .