When we have no hero, we will definitely want to hire one.
We will also want to hire heroes who already pay for more than themselves by coming with an army that has more value than the hero costs.
Allowed the AI to cast spells that are aimed at a location instead of a unit. For example meteor shower.
Fixed an issue that caused the AI to not kill unit-stacks with spells due to only considering stacks where at least one unit survives in it's score-calculations.
Fixed a bug that prevented the AI from using spells when attacking an enemy settlement that has towers.
The bug was caused by noticing how greatly effective spells would be against towers but not being able to actually target them.
By skipping invalid targets, this no longer is an issue.
Due to morale-considerations the AI sometimes calculated that their strongest army after doing an exchange had slightly lower total value than the army they used before.
But by using unsigned "slightly lower" became near infinite.
So they constantly wanted to upgrade their army because they considered it more useful than anything else.
Changing the unsigned into signed fixes this.
It is now possible to switch to an AI-variant that uses hand-written heuristics for decision-making rather than the FuzzyLite-engine. This is configurable in nkai-settings.json via the new parameter "useFuzzy".
The information of whether objects like a redwood-observatory or subterranian gates have been interacted with by the AI will now be retrieved from the game-state instead of using an AI-internal memory that won't survive loading a save-game.
Nullkiller suggested that this change would help to further fix inconsistent behavior by the AI. I tested it and it did indeed fix different orders of how AI does things.
"Important to make count 1 to not relay on object addresses
They are source of random" - Nullkiller
When I was playing a very large map, computer's turn took up to 2 minutes and that give me time to fire up perf and profile the code while waiting.
I noticed that 1.7% of time was spent in function CPathfinderHelper::getNeighbors and half of that on a single load from an array backing multi_array of TerrainTiles. That signals the CMap::terrain is too big to fit in CPU caches.
This patch reorders fields in TerrainTile struct, shrinking it from 96 bytes to 80 bytes and hopefully helping to keep more tiles in the cache and speeding things up a little bit.
