Hey everyone,
Last time I said the next dev log would be about the match engine, specifically how the AI plays save rounds. This is that one. There's also a Discord server now. Details at the bottom.
You've lost the round. Two of you are alive, the bomb is down, the timer isn't
going to save you. So you don't trade a rifle for a fight you can't win. You
break off, you stay alive, and you start the next round with a gun instead of a
pistol. Done well, that's one of the quiet skills that separates a good team from a bad
one. Done badly, it looks like someone wandering into a corner and dying with a
full inventory. FRAGLINE was doing the second one. The report that kicked this off, more or
less word for word: "the saver runs off into the distance and won't even
react to the guy rushing him, and he always saves in spawn." Both halves were true, and pulling on that thread took most of the month. <
Every player in the game has a field of view, and what he can see depends on
where he's facing. A player who is standing still faces wherever he's aiming. The saving player wasn't aiming at anything. He was told to run, and while he
was running the code that decides where he looks skipped him entirely. So his
face stayed pointed the way he happened to be moving, and once he arrived, it
froze there. His field of view went with it. A hunter walking up behind him was
outside the cone, which meant, as far as the engine was concerned, invisible. He wasn't ignoring the guy rushing him. He genuinely couldn't see him. Sound had the same problem from a different direction. Footsteps and gunfire
were reaching him fine, the information was sitting right there, but every
piece of code that reads it started with "if he's saving, never mind" and
turned around. Deaf on top of blind. And the decision itself was one way. Once a player entered save mode, nothing
in the round could take him out of it. Not an enemy walking into his lap, not
his teammate winning the fight behind him.
That was the part I got wrong for the longest time, and it took someone playing
the game to spot it rather than me reading code. The engine had a list of hand marked save positions per map, and those all sit
near the spawns. When none of them was close, it invented one by looking in a
ring around the player. Both of those are the same instinct: stay near where
you already are. So a CT who decided to save never left his own half of the
map. Not rarely. Never, on two of the three maps I test on. His save positions
collapsed into four to six spots per map, the same corners every time. That isn't a player protecting a rifle. That's a player who only uses the map
to walk through. Save positions are now drawn from the whole map. The engine already builds a
network of navigation nodes for another system, and the saver picks from that:
anywhere he can actually reach, scored on how far it is from where the enemy is
known to be, whether the route there walks him through them, whether he can be
seen from the threat, and how many ways in the spot has. Not the furthest point
and not the nearest one, the one that survives all four questions. Now CT save positions spread across twenty to twenty four distinct areas per
map, and roughly a third of the time he crosses into the other half of the map
to do it. Which is what you'd do.
A CT who saves after the bomb is planted and a CT who saves before it are doing
two different jobs, and the engine was treating them identically. After the plant, the site is a countdown. He gets out, and out means far: he
now ends up 60 to 85 units from the bomb, and never inside the blast radius. He
used to settle 33 units away, which on a bad day is close enough to be killed
by the explosion he ran from. Zero positions inside the blast, out of every
planted round I measured. Before the plant, he doesn't have to abandon anything. The round isn't
decided yet, his site might still matter, and a short reposition near what he
was already holding is a perfectly good save. So he stays close, around 15
units, instead of sprinting across the map for no reason. The T side works the same way now. The question that picks the profile is the
same on both sides, where is the bomb, but the sign is flipped: a CT is running
from the bomb, a T is running from having lost it. If it's still with his team
the round isn't over, so he stays near the man carrying it. If it's on the floor
because the carrier died, the objective is gone and he gets out. The T side also calls the save together now, and on a different trigger. A CT
team gives up on the round when the defuse can't happen in time. A T team gives
up when the round is arithmetically lost: the bomb is on the floor with a CT
closer to it than anyone left, or there isn't enough clock to reach a site and
plant. Being down a man isn't enough on its own. If they're even or ahead they
go and take the bomb back, which is what you'd do.
Even with a good position, the saver was a statue. He'd arrive, lock onto one
direction and stand there. And because the direction came from a rough guess at
where the threat was, that direction was frequently a wall. Forty percent of
the time he spent posted, he was aiming at solid geometry. He holds an area now, not a point. The engine works out which ways into that
area are actually open, and those are what he watches: rotating between them
when he has no information, swinging to the one nearest a noise when he does.
Before the aim gets written, it's checked against the map from where he's
actually standing, so if the angle has drifted into a wall it gets thrown out
and recomputed. Wall time went from forty percent to seven. Small thing to describe. It's the difference between a man hiding and a man
holding an angle.
The first one I never would have found by watching. The engine can replay any
round from its seed and print every decision with the reason behind it, and
that's where this turned up. Saving is supposed to be off the table when you're up a man. Five versus four,
the odds of anyone saving are zero. Except the AWP bonus and the retake call
were both being added on top of that zero instead of scaling it, so a player
holding an AWP had a 15% chance of abandoning a round he was winning. It had
been in there for a long time, hiding in exactly the kind of round nobody
reviews because you won it anyway. The second: the CT hunting a saver was reading the saver's true position
through walls. The T side had been fixed a while back and the CT side was
simply never revisited. Now he sweeps the places a saver would plausibly be,
and sometimes he sweeps the wrong one and the save works. Which is the point. While I was in there, two more rules landed. Deep man disadvantage, five versus
one or five versus two, is now an automatic save rather than a dice roll, and
the call is the in game leader's rather than five players deciding
independently. A team used to split, half of them saving while the other half
fought.
Three maps, eight seeds each, every change measured before and after on the
same fixed seeds.
Rifles actually saved: 95 to 100%. The player who decides to save is
alive at the end of the round. This is the number that matters and it's the one
I wasn't measuring at first. I spent a week optimising "did he reach his
position", which turned out to be the wrong question.
Distance to the nearest enemy once he settles: 20 units to 50. He's
not loitering next to the people he ran from any more.
CT save positions per map: 4-6 areas to 20-24. The map is being used
instead of walked through.
Time spent aiming at a wall: 39% to 7%.
Positions inside the blast radius after a plant: zero.
And the one I watch hardest: the T side round win rate. It was 52% when this
started and it's 50% now. Every change here makes savers survive more often,
which feeds the next round's economy, and there was a point mid month where that
had pushed the balance eleven points in the T side's favour. Tightening the
position rules pulled most of it back, and the fix in the next section took the
rest. Same matches, same seeds, before and after.
I wrote above that a spot gets scored on how many ways in it has. That score
asks a plain question: standing here, in how many directions can you see
anything? It looks in twelve directions at a fixed distance and counts the ones
that aren't blocked. The distance was twelve units, and that was too far. At twelve units nearly
every direction runs into something, on every map I have. So nearly every
candidate came back with no open ways at all, which the engine reads as a blind
hole and throws out. Nothing ever reached the ideal, the score stopped
separating anything, and the choice quietly fell through to the other terms.
The measure had been switched off for months without saying so. I had blamed Overpass for this. Wide open map, I said, you can't tell a
sheltered corner from open ground there. Measuring it properly said the exact
opposite: Overpass has the most enclosed save positions of any map I test on.
The longest sight line out of a typical spot there is about twelve units. On
Dust it's over twenty. The map was never the problem. The ruler was. Two things changed. The radius came down to six units, which is close enough
that the count means something again. And counting ways in was never the whole
question, so there's a second measure now: how long is the longest line of sight
out of this spot. Two corners can each have three ways in, but if one of those
ways is a forty unit corridor, that corner is AWP food. A spot on a long line
gets scored down, and past a hard limit it gets thrown out the same way a blind
hole does. Blind holes are now at zero on every map, and outside Dust almost nothing sits
on a long sight line any more.
It's up, it's open, and it's the easiest place to reach me: Join Discord Server What it's for: bug reports, saves that went strange, feature requests, and the
build-by-build stuff that's too small for an announcement but that you'd
probably want to know about. If you've already sent me something through the
Steam forums, thank you, keep doing that too. This is just a faster line. Everything in this post came from one player describing what looked wrong to
him. That's worth more than it sounds.
Early Access, first. Save rounds are in a state I'm happy with now, so the next
pass goes somewhere else in the match engine. One loose end stays on the list:
on Dust, a T who saves still ends up on a long sight line about a fifth of the
time, because the sheltered corners on that map are the ones he's walking away
from.
FRAGLINE is a solo project. Every bug report and suggestion so far has gone
straight into the build, so keep them coming. Early Access: August 26. The demo is live now.
Wishlist it on the store page so you know when it's out. FRAGLINE