I'm working on an isometric 2D game with moderate-scale multiplayer, approximately 20-30 players connected at once to a persistent server. I've had some difficulty getting a good movement prediction implementation in place.

Physics/Movement

The game doesn't have a true physics implementation, but uses the basic principles to implement movement. Rather than continually polling input, state changes (ie/ mouse down/up/move events) are used to change the state of the character entity the player is controlling. The player's direction (ie/ north-east) is combined with a constant speed and turned into a true 3D vector - the entity's velocity.

In the main game loop, "Update" is called before "Draw". The update logic triggers a "physics update task" that tracks all entities with a non-zero velocity uses very basic integration to change the entities position. For example: entity.Position += entity.Velocity.Scale(ElapsedTime.Seconds) (where "Seconds" is a floating point value, but the same approach would work for millisecond integer values).

The key point is that no interpolation is used for movement - the rudimentary physics engine has no concept of a "previous state" or "current state", only a position and velocity.

State Change and Update Packets

When the velocity of the character entity the player is controlling changes, a "move avatar" packet is sent to the server containing the entity's action type (stand, walk, run), direction (north-east), and current position. This is different from how 3D first person games work. In a 3D game the velocity (direction) can change frame to frame as the player moves around. Sending every state change would effectively transmit a packet per frame, which would be too expensive. Instead, 3D games seem to ignore state changes and send "state update" packets on a fixed interval - say, every 80-150ms.

Since speed and direction updates occur much less frequently in my game, I can get away with sending every state change. Although all of the physics simulations occur at the same speed and are deterministic, latency is still an issue. For that reason, I send out routine position update packets (similar to a 3D game) but much less frequently - right now every 250ms, but I suspect with good prediction I can easily boost it towards 500ms. The biggest problem is that I've now deviated from the norm - all other documentation, guides, and samples online send routine updates and interpolate between the two states. It seems incompatible with my architecture, and I need to come up with a better movement prediction algorithm that is closer to a (very basic) "networked physics" architecture.

The server then receives the packet and determines the players speed from it's movement type based on a script (Is the player able to run? Get the player's running speed). Once it has the speed, it combines it with the direction to get a vector - the entity's velocity. Some cheat detection and basic validation occurs, and the entity on the server side is updated with the current velocity, direction, and position. Basic throttling is also performed to prevent players from flooding the server with movement requests.

After updating its own entity, the server broadcasts an "avatar position update" packet to all other players within range. The position update packet is used to update the client side physics simulations (world state) of the remote clients and perform prediction and lag compensation.

Prediction and Lag Compensation

As mentioned above, clients are authoritative for their own position. Except in cases of cheating or anomalies, the client's avatar will never be repositioned by the server. No extrapolation ("move now and correct later") is required for the client's avatar - what the player sees is correct. However, some sort of extrapolation or interpolation is required for all remote entities that are moving. Some sort of prediction and/or lag-compensation is clearly required within the client's local simulation / physics engine.

Problems

I've been struggling with various algorithms, and have a number of questions and problems:

1. Should I be extrapolating, interpolating, or both? My "gut feeling" is that I should be using pure extrapolation based on velocity. State change is received by the client, client computes a "predicted" velocity that compensates for lag, and the regular physics system does the rest. However, it feels at odds to all other sample code and articles - they all seem to store a number of states and perform interpolation without a physics engine.

2. When a packet arrives, I've tried interpolating the packet's position with the packet's velocity over a fixed time period (say, 200ms). I then take the difference between the interpolated position and the current "error" position to compute a new vector and place that on the entity instead of the velocity that was sent. However, the assumption is that another packet will arrive in that time interval, and it's incredibly difficult to "guess" when the next packet will arrive - especially since they don't all arrive on fixed intervals (ie/ state changes as well). Is the concept fundamentally flawed, or is it correct but needs some fixes / adjustments?

3. What happens when a remote player stops? I can immediately stop the entity, but it will be positioned in the "wrong" spot until it moves again. If I estimate a vector or try to interpolate, I have an issue because I don't store the previous state - the physics engine has no way to say "you need to stop after you reach position X". It simply understands a velocity, nothing more complex. I'm reluctant to add the "packet movement state" information to the entities or physics engine, since it violates basic design principles and bleeds network code across the rest of the game engine.

4. What should happen when entities collide? There are three scenarios - the controlling player collides locally, two entities collide on the server during a position update, or a remote entity update collides on the local client. In all cases I'm uncertain how to handle the collision - aside from cheating, both states are "correct" but at different time periods. In the case of a remote entity it doesn't make sense to draw it walking through a wall, so I perform collision detection on the local client and cause it to "stop". Based on point #2 above, I might compute a "corrected vector" that continually tries to move the entity "through the wall" which will never succeed - the remote avatar is stuck there until the error gets too high and it "snaps" into position. How do games work around this?