Five military flavors for operations and tactical AI

[CIVITAS-John]

1. Mod version

5.4.6

2. Installed Components

Vox Populi with EUI

3. Additional mods used

N/A

4. Describe the Issue

Five military flavors for operations and Tactical AI

Hi folks,

This is a proposal to add five flavors into VP, gated by a modmod flag (or not?), to give grand/military/economic strategies more room to steer campaign-level (i.e., operation and tactical zone) and tactical-level (combat simulator) AI. It does not create any AI impact on its own, but rather creates room for modmods to experiment with them. For example, we can make a civilization under EmpireDefenseCritical (i.e., the homeland is significantly under threat) a major bonus on choosing defensive tactics and on holding cities, while reducing their tendency to launch any (likely fruitless) attack. Or we can make a nearly bankrupt civilization a major debuff on sacrificing units, as that only adds to their economic burden.

Below is an AI-written proposal (with my light revisions).

The five flavors

Flavor	Higher value	Lower value
FLAVOR_RISK	Prefer caution and avoid losses	Accept losses and exposure to press an advantage
FLAVOR_OCCUPATION	Get cities	Leave cities for other objectives
FLAVOR_ATTRITION	Kill enemy units	Give less priority to killing enemy units
FLAVOR_HOLD_CITY	Hold cities	Yield cities
FLAVOR_HOLD_GROUND	Hold territory and field position, resisting yielded ground	Yield territory when useful

The implemented capture baseline uses a fixed u8[5] array in the order RISK, OCCUPATION, ATTRITION, HOLD_CITY, HOLD_GROUND. Values range from 0 through 100, with 50 neutral. Risk is captured as the inverse of clamped stock offense, 100 - 10 * clamp(FLAVOR_OFFENSE, 0, 10); the other four slots are 50. The campaign carries a global vector and an empty masked per-civ override section, and each request carries its final vector. New tactical scoring remains future work.

Future values use signed contributions from active strategies, through the same mechanism as AIEconomicStrategy_Player_Flavors and AIMilitaryStrategy_Player_Flavors. Preserve whether risk was assigned, including an explicit zero, then resolve the contributions into the final 0 through 100 scale with 50 neutral. Explicit-versus-unset risk resolution and the other strategy producers remain future work.

How the values flow

Currently, with one exception (Offense), military AI is not impacted by flavors. We propose to add the flavors into the following subsystems. The baseline preserves existing behavior while the future producers and consumers are defined.

flowchart LR
  GS[Grand strategy<br/>flavor modifiers] --> SUM
  ES[Active economic strategies<br/>Player_Flavors] --> SUM
  MS[Active military strategies<br/>Player_Flavors] --> SUM
  LP[Leader personality] -. excluded .-> SUM
  CS[City strategies] -. excluded .-> SUM
  SUM[Future strategy contributions]
  SUM --> OPS[Operations<br/>CvMilitaryAI]
  SUM --> POST[Zone posture<br/>CvTacticalDominanceZone]
  SUM --> RES[DLL tactical intent resolution]
  STOCK[Stock offense] -->|baseline risk capture| RES
  RES --> SIM[Combat simulator<br/>TacticalAIHelpers]

Loading

Proposed flavors tilt discretionary decisions

Operations (CvMilitaryAI::UpdateOperations)

flowchart LR
  A[RISK] --> D1[Start an offensive?<br/>DoCityAttacks reserve and war-state gate]
  A --> D4[Abort on losses<br/>UnitWasRemoved half-slot rule<br/>offensive ops only]
  A --> D5[New: abandon an unpromising offensive<br/>next to the mechanical aborts]
  O[OCCUPATION] --> D2[City attack vs pillage<br/>RequestCityAttack / RequestPillageAttack]
  O --> D3[Target ranking<br/>ScoreAttackTarget]
  H[HOLD_CITY] --> D3

Loading

Zone posture (SelectPostureSingleZone, then SelectPostureMultiZone)

Posture is currently chosen from territory type, overall and ranged dominance, melee balance, counterattack danger, and whether the zone city is in danger of falling. Flavors shift the thresholds for each stance. Below are AI generated examples, and I think more (programmatic) tweaking might be necessary.

flowchart LR
  A[RISK] --> P1[caution ↔ hold position ↔ accept exposure<br/>withdraw over counterattack]
  O[OCCUPATION] --> P2[surgical city strike]
  H[HOLD_CITY] --> P3[hold cities and resist withdraw<br/>incl. multi-zone 'support a neighbor' override]
  F[HOLD_GROUND] --> P4[resistance to yielding territory and field position]
  T[ATTRITION] -. future enemy-unit destruction posture .-> P0[ ]

Loading

Combat simulator (TacticalAIHelpers::FindBestUnitAssignments)

All five values are passed in explicitly, as the simulator does not read global flavor state.

flowchart LR
  A[RISK] --> S1[unit-loss threshold, min HP to simulate,<br/>captured baseline preserves stock offense behavior]
  O[OCCUPATION] --> S2[city damage and capture weights]
  T[ATTRITION] --> S3[unit damage and kill weights]
  H[HOLD_CITY] --> S4[value of holding cities]
  F[HOLD_GROUND] --> S5[value of holding territory and field position]

Loading

Selected risk baseline

The baseline captures risk before calling the tactical simulator. It derives the existing loss and minimum-health thresholds from the stock FLAVOR_OFFENSE value through the inverse risk formula, preserving current behavior. Future explicit risk resolution and new scoring are deferred.

The other four baseline slots are neutral at 50. Future per-civ overrides retain presence masks so individual slots can inherit global values.

Minor Notes

• Consider deterministic A* tie-breaking in CvAStarNode.h. Preserve total-cost and heuristic priorities, then define stable coordinate and stop-layer ordering shared by the DLL and simulator. Otherwise, the offline simulator cannot reproduce exact behaviors.

[Hokath]

If the Aggression flavor has strong control over behavior it would be an excellent handle for tuning AI decision making.

Amongst Leader flavors, we set values for

'FLAVOR_MILITARY_TRAINING', 'FLAVOR_OFFENSE', 'FLAVOR_DEFENSE', 'FLAVOR_CITY_DEFENSE', 'FLAVOR_EXPANSION'

where we do
Conquerors, Coalitionists, Diplomats, and Expansionists
as templates we then edit further.
Would it make sense to me that the aggression baseline would take into account this Military Training value?
(8, 6, 3, 5 for the templates above)
Since it represents somehow the Leader's willingness to divert economy to unit training (and some Civs have special bonuses there, either extra XP to starting units, or +% production to building new units, purchasing, etc.)

[CIVITAS-John]

@Hokath The goal is to have all five flavors control AI tactical/operational behaviors.

Aggression is currently Offense, I just propose to split it from the original flavor which also influences promotion/unit build choices. So Offense = I want to build an army for offense, Aggression = I want my military to fight aggressively right now.

[CIVITAS-John]

Updated the five flavors!