|
buddye -> RE: BOBII 2.08 AI performance (6/8/2008 9:35:04 PM)
|
Hi Guys, just a short version of some AI performance notes so you can better understand 2.08 for anyone that may be interested.
The 2.08 AI Performance Story
The redesign of the AI Maneuver Selection
The redesign of the AI maneuver Selection Criteria was driven by the need to become more deterministic and less random in selecting AI maneuvers. I felt the need to move in this direction to improve the AI offence and defense so the AI selection software had more control. Of course there is a big danger in this strategy. We do not want BOBII to become repeatable. This will become a fine line to follow in the future and I will need both our testers and customers help and feedback.
Here is a summary of the new selection criteria design:
1. Altitude (how much altitude and rate of change)
2. Speed (how much speed, and rate of change)
3. Position of A/C to each other (none to tail, tail to tail, nose to beam, tail to beam or left, right, front, back)
To implement the new selection criteria I built a 3 X 6 matrix (a truth table) and many new programs. For each A/C (the unfriendly and the AI or the player), I designed programs to look at Altitude, Speed, and position and to first try and select the best maneuver option (aggressive or defense) based on AI skill level.
Altitude and speed can also be thought of as “energy” as altitude can be turned into speed.
The new design uses the A/C’s current position but I am thinking about implementing the A/C’s “lead” position (his future position which would be a small delta in the future).
Anyway, this new design is a long term work BOBII AI strategy (work in progress) that I can work on as I have new ideas for the AI from our testers and customers. It should position the BOBII AI design for future improvements by using more deterministic rather than random selection criteria. The goal is very simple "to improve the AI performance" and keep BOBII the best off line AI.
My gut feel based on my experience and testing is that the AI is now stronger at selecting the best/correct maneuver, avoiding bad positions (like low altitude), better at avoiding low energy, and in general a bit stronger fighter and defender. Of course, the Player will always “win” with experience/practice but if we give the player a better fight (scrap), I personally consider that real progress.
BTW, I also gave the Terminator a bit of boost in performance and tune-up so you might want to try it out. With the boost in Terminator performance, he was getting “cocky” so I implemented a spinout feature. The Terminator flies so close to the edge now that he has a tendency to “spinout”. This is when you can get him.
I also implemented a new AI feature called “Flying Factor (FF)”. This is the knowledge of the AI pilot to fly a given maneuver (experience) and how well the AI pilot will actually fly the given maneuver (skill). The FF is based on the Skill Level (customer selected in Instant Action Missions and software assigned in the Campaign). The Terminator is assigned a Skill Level of Hero2 (the highest in the game) so that is where he gets his boost in performance (edge).
The AI Performance Dependencies
Skill Level
The AI performance is dependent on AI Skill Level (which is customer selected in Instant Action and SW assigned in the campaign). BOBII AI do make mistakes (spin, crash, dumb maneuvers, bad judgment, shoot late , shoot bad, etc) which is after all very human.
The skill level of the AI is key in making decisions on about everything with respect to AI performance like (1) how well the AI fly, shoot, shoot fast, slow, accuracy, or not shoot, (2) how well the AI fly, what maneuvers are selected, and how well the AI will fly the chosen maneuver.
Random Numbers (Luck of the draw)
The BOBII AI A/C is also dependent on luck (specifically on random numbers). Random number decisions are coded through out the AI code. BOBII’s random approach keeps BOBII from doing the same thing each time. Even something as simple as the direction to start a maneuver (left or right), I will use a random number to decide (why hard code something when you can use a random number).
For example, most BOBII vertical maneuvers use a random number to assign a length of time to for a specific vertical maneuver (Like Zoom). The AI pilot will sometimes cut off early, or late, or somewhere in the middle. If early the maneuver may carry too much speed and if late the AI may slow down so much that control is lost (very human).
The bottom line is that the customer will always see a somewhat different maneuver (very good, good, not so good, and loss of control) both because of the random implementation and the different physic’s parameters (speed, roll, heading, pitch, and overall energy) going into each maneuver.
The very real downside of random numbers is it is very hard to test (not repeatable) and the processing power used.
Conditions
The conditions for each maneuver are always different (energy, speed, altitude, skill, damage, and enemy position). This also changes how the maneuver is performed. A damaged AI will not fly as well as an undamaged AI.
Maneuver Selection
The first key decision to be made is to select either an aggressive or defensive maneuver. This is a complex decision based on the available information on both the Player and the Enemy AI or the friendly AI and the enemy AI. The data considered for both Player and enemy is speed, altitude, and the position of the targeting AI and the AI being targeted with respect to each other.
After selecting either an aggressive or defensive, then a random approach is used to select a category (choose good , choose bad , or choose neither good nor bad maneuver).
Maneuvers are then divided into three parts Vertical, Horizontal, and dive for each of our categories (choose good, choose bad, and choose neither good nor bad maneuver).
The individual maneuver selection is then based on speed, altitude, and position of both the player and the enemy AI.
|
|
|
|
