AI Recruitment

A new recruitment candidate action (CA) was introduced and made the default in Wesnoth 1.11. It is much more configurable than previous versions and supports multiple leader recruiting. This page describes how it works and the parameters available for customizing it.

How it works

Map analysis

In a first step, the CA analyzes the map and tries to find "important hexes". Those are locations on the map where a fight will occur with high probability. The important hexes are later used in combat simulations to calculate an average defense rating for the combatants.

When the game is running in debug mode and the game is started with the parameter --log-info=ai/recruitment the important hexes are marked with white 'x' markers on the map.

Score Map

A score map is created for all leaders who are able to recruit. All unit types from the recruit and recall lists are mapped to a score.

In the end, the scores represent the desired unit ratios on the map. This means the AI will not just recruit the unit with the highest score, but in such a way that the mix of units on the map comes as close as possible to the mix represented by the scores. The scores are filled with values coming from combat analyses (see below).

After that the AI modifies the scores in several ways:

• Similar units get a penalty for being similar. Similar units are units in one advancement tree.

Example (Archer can advance to Ranger):
			before	after
Elvish Fighter: 	  50	 50
Elvish Archer:		  50	 25
Elvish Ranger:		  50	 25

• The aspects recruitment_more, recruitment_diversity and recruitment_randomness are handled (see below).

Combat Analysis

For each unit type the AI leaders can recruit, the CA simulates fights against all enemy units on the map. If there are less then 5 enemy units, the enemy recruitment list(s) are also taken into account.

After combat analysis the score map may look like this:

Goblin Spearman score:		16.1252
Naga Fighter score:		0
Orcish Archer score:		61.6442
Orcish Assassin score:		88.0737
Orcish Grunt score:		100
Troll Whelp score:		26.7646
Wolf Rider score:		0

Implementation details for interested readers

The combat analysis uses the function compare_unit_types(A, B). It takes two unit types and simulates a fight, using a cache because simulation is expensive. The function returns a positive value if unit type A is better than B and a negative value if unit type B is better than A. If the return value is 2.0, it means that unit type A is twice as good as unit type B.

Here is a important code snippet from compare_unit_types():

double value_of_a = damage_to_b / (b_max_hp * a_cost);
double value_of_b = damage_to_a / (a_max_hp * b_cost);

if (value_of_a > value_of_b) {
  return value_of_a / value_of_b;
} else if (value_of_a < value_of_b) {
  return -value_of_b / value_of_a;
} else {
  return 0.;
}

damage_to_b comes from simulations and is the average damage dealt by A when A attacks (A chooses weapon) plus the average damage dealt by A when B attacks (B chooses weapon). damage_to_a is equivalent.

The return value of compare_unit_types() is multiplied by the enemy's current hitpoints and added to the score map.

Included in the damage calculations are average defense (coming from "important hexes"), resistance, resistance abilities, average time of day, drain, poison, berserk, swarm and slows.

At this point the scores can be quite inconvenient for a generalized treatment. They could all be negative, very high or too similar. Thus, the scores are transformed linearly. The resulting scores are numbers between 0.0 and 100.0. The minimum depends on the aspect recruitment_diversity.

Simplified code snippet:

double new_100 = max_score;
double new_0 = max_score - (recruitment_diversity * (max_score - average_score));

BOOST_FOREACH(double& score, scores) {
  score = 100 * ((score - new_0) / (new_100 - new_0)); // linear transformation
  if (score < 0.) {
    score = 0.;
  }
}

Aspect recruitment_instructions

This is a powerful aspect for controlling recruitment via WML. Its syntax is quite complex and requires the composite aspect configuration. At its core are the [recruit] tags, which are the instructions to the candidate action defining which units are to be recruited.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
       turns=
       [value]
         #(here multiple [recruit] and/or [limit]-tags)
         [recruit]
           #(attributes)
         [/recruit]
         [limit]
           #(attributes)
         [/limit]          
       [value]
    [/facet]
  [/aspect]
[/ai]

Attributes inside [facet]:

• turns="": (string) This key takes a comma separated list containing numbers specifying the turns. '-' can be used between two values to define a range. An empty string means all turns.
• The other keys possible in [facet] tags can, in principle, also be used, but note that invalidate_on_gamestate_change and invalidate_on_minor_gamestate_change have no effect on [recruit] tags. They are read at the beginning of the turn and are immutable for the remainder of the turn.

Attributes inside [recruit]:

• type="": (string) This key takes a comma separated list containing unit types, usages or levels. Common values for usage are 'scout', 'fighter', 'archer', 'healer' and 'mixed fighter'. An empty string means 'all units'. If more then one unit is specified, the AI decides according to the score map what to recruit.
• number=-1: (integer) A number n greater than 0 instructs the AI to recruit n units. '-1' means 'as much as possible'. '0' means 'do not recruit'.
• importance=1: (integer) Priority of the recruitment. The AI first recruits units with the highest importance, until the specified limits are reached. If gold is lacking or the castle is full, only the most important units are recruited, all other [recruit] jobs are dropped.
• leader_id="": (string) id of the leader to execute this recruitment. Empty string means 'all leaders'. Note that this is only a recommendation to the AI. If the specified leader has no free hexes available, the AI uses other leaders to do the job.
• total=no: (boolean) If set to 'yes', the AI counts the own units on the map which match at least one of the given types and then recruits the difference between number and the counted amount.
• blocker=yes: (boolean) If set to 'yes', the AI stops recruiting if this job cannot be done (for example, because of lack of gold or a [limit]). If set to 'no', the AI drops this job and tries to continue with less important ones.
• pattern=no: (boolean) If set to 'yes', the unit to recruit is not chosen by the AI, but randomly according to the frequency in the type attribute. For example, when type=Orcish Grunt, Orcish Grunt, scout and number=6, the AI recruits 6 units, with the probability that one unit is a grunt being twice as high as that of a scout. If the type attribute is empty, the AI recruits randomly. (See also recruitment_pattern)

Attributes inside [limit]:

• type="": (string) This key takes a comma separated list containing unit types, usages or levels. An empty string means 'all units'.
• max=0: (int) The maximum of units of the given type(s).

Notes:

• [limit] has higher priority than [recruit].
• The recruitment_save_gold aspect has higher priority than [recruit]. So if exact recruiting is wished, deactivate recruitment_save_gold (see below).
• If recruitment is specified for a turn, the AI does not recruit any other unit by default. (So if the AI is told to recruit 1 scout in turn 1, then the AI will only recruit 1 scout and no other units). To prevent this behavior one can add this [recruit] tag:

[recruit]
   importance=0
[/recruit]

The AI then falls back to normal recruiting when all other [recruit] tags are dealt with. This is also the content of the [default] facet for the aspect.

• Even if only a [limit] tag is used in the aspect declaration, the [default] facet will get overwritten. Include the [recruit] tag mentioned in the previous point to let the AI recruit something.
• Only one facet at a time can be active. When there are more [facet]s defined in a turn, the behavior is undefined. Use several [recruit] and/or [limit] tags in the same facet instead.
• Do not define more than one [recruit] tag with the same importance. Also note that the aspect recruitment_pattern expands into a [recruit] tag with importance=1.

Examples

Recruit 3 Grunts (and nothing more) in turn 3-5.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      turns=3-5
      [value] 
        [recruit]
          type=Orcish Grunt
          number=3
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Recruit as many scouts as necessary until there are 4 scouts in total and then recruit other units.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      [value] 
        [recruit]
          type=scout
          number=4
          total=yes
          importance=1
        [/recruit]
        [recruit]
          importance=0
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Recruit 6 Grunts or level 2 units (whatever seems better for the AI) and nothing more.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      [value] 
        [recruit]
          type=Orcish Grunt, 2
          number=6
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Recruit 5 scouts with leader1 and 5 Grunts with leader2. Do even try to recruit the Grunts if leader1 can not recruit all of the scouts.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      [value] 
        [recruit]
          type=scout
          number=5
          importance=2
          leader_id=leader1
          blocker=no
        [/recruit]
        [recruit]
          type=Orcish Grunt
          number=5
          importance=1
          leader_id=leader2
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Recruit random units always.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      [value] 
        [recruit]
          type=
          pattern=yes
          importance=1
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Do not recruit in turn 4.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      turns=4
      [value] 
        [recruit]
          importance=1
          number=0
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

Do not have more than 6 units on the map.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      [value] 
        [limit]
          type=    #empty type means all units
          max=6
        [/limit]
        [recruit]         #Do not forget to include the default
          importance=0    #[recruit]-tag. Otherwise the AI won't
        [/recruit]        #recruit anything.
      [/value]
    [/facet]
  [/aspect]
[/ai]

This will not work because only one [facet] can be active.

[ai]
  [aspect]
    id=recruitment_instructions
    [facet]
      turns=1-10
      [value] 
        [recruit]
          type=scout
          number=5
          importance=1
        [/recruit]
      [/value]
    [/facet]
    [facet]
      turns=3
      [value] 
        [recruit]
          type=Orcish Grunt
          number=3
          importance=2
        [/recruit]
      [/value]
    [/facet]
  [/aspect]
[/ai]

To achieve this behavior one has to create a [facet] for turn 1-2, a [facet] for turn 3 and another [facet] for turn 4-10.

Aspect recruitment_save_gold

How it works

For several reasons it can be a good idea not to spend all gold at once:

• To save the upkeep,
• To wait for the enemy to recruit first to counter easily,
• To save gold for the next scenario when the enemy is almost defeated.

At the same time it is a good idea (especially for the AI) to recruit in 'waves' (not to recruit one unit each turn).

For this there is the aspect recruitment_save_gold.

The Recruitment CA is always in one of the following states:

• NORMAL,
• SAVE_GOLD,
• SPEND_ALL_GOLD,
• LEADER_IN_DANGER.

Except the state LEADER_IN_DANGER, the states are persistent over turns.

The AI keeps track of a 'unit_ratio'. This is our_total_unit_costs / enemy_total_unit_costs whereas the costs are the sum of the cost of all units on the map weighted by their HP. Also the AI tries to estimate the total income over the next 5 turns. (This is a quite rough estimation, but that's fine.)

When the AI is in state NORMAL and the unit_ratio exceeds a predefined threshold ('begin') and the estimated income is positive, the AI will switch to the state SAVE_GOLD. Now the AI does not recruit units anymore until the unit_ratio falls below another predefined threshold ('end'). Then the AI will switch to the NORMAL state and recruit again.

With this behavior the AI can save quite a lot of gold. When the AI has more gold then a predefined value ('spend_all_gold'), it will switch into the state SPEND_ALL_GOLD and recruit a big army to start an offensive wave.

The AI will ignore all those gold saving strategies when there is a LEADER_IN_DANGER. This happens when an enemy is near the leader (5 hexes).

Important note: Although the intuition says that the AI will play better when it tries to save some gold, it doesn't. Tests shown that it is always good for the AI to have as many units as possible. (Probably it will do better in village-capturing then). The aspect is activated by default because playing against wave-like recruiting is fun.

The Aspect

This aspect requires also the composite form.

[ai]
  [aspect]
    id=recruitment_save_gold
    [facet]
      [value]
        #(attributes)
      [/value]
    [/facet]
  [/aspect]
[/ai]

Attributes inside [value]: (with default values)

• active=2: (int) From this turn on the aspect will be active. So the AI will always spend all gold until the defined turn. active=0 can be used to always deactivate gold saving. (The default value is 2 so nobody will get confused if the AI doesn't recruit in the first turn)
• begin=1.5: (double) See explanation above.
• end=1.1: (double) See explanation above.
• spend_all_gold=-1: (int) See explanation above. If set to -1 the AI will never switch to the state SPEND_ALL_GOLD.
• save_on_negative_income=no: (boolean) If this is set to yes, the AI will switch to the state SAVE_GOLD even if the income is negative. (estimation over 5 turns, village gain and unit loss estimation included).

Example

This example is filled with all default values and can be used for quick copying.

[ai]
  [aspect]
    id=recruitment_save_gold
    [facet]
      [value]
        active=2
        begin=1.5
        end=1.1
        spend_all_gold=-1
        save_on_negative_income=no
      [/value]
    [/facet]
  [/aspect]
[/ai]

A macro can be used to deactivate recruitment_save_gold:

{AI_ASPECT recruitment_save_gold {AI_DEACTIVATE_SAVE_GOLD} }

Other Recruitment Related Aspects

• recruitment_diversity=2.0 (double) When this value is high, the AI will recruit more units which are currently rare on the map. When the value is 0.0, the AI will only recruit the best unit.(see above for implementation details).
• recruitment_more="" (string) This key takes a comma separated list containing unit-types, usages or levels. This is meant to let a scenario editor make an easy hack when he/she wants the AI to recruit more units of a specific type. (25 will be added to the unit-type's score.) With recruitment-more="Orcish Grunt, Orcish Grunt" it is possible to add 50 to the Grunt's score.
• recruitment_randomness=50 (int) To each score a random value between 0 and recruitment_randomness will be added. Don't hesitate to use a high value (like 200) to increase randomness.
• recruitment_pattern="" (string) This key takes a comma separated list containing unit-types, usages or levels. Common usages are: 'scout', 'fighter', 'archer', 'healer' and 'mixed fighter'. This tells the AI with what probability it should recruit different types of units. The usage is listed in the unit type config files (see data/core/units/ for mainline units; see also UnitTypeWML).
  • For example, recruitment_pattern=fighter,fighter,2 means that the AI recruits on average twice as many fighters as level 2 units. It does not mean that it recruits two fighters first, then a level 2 unit, then two fighters again, etc.
  • Internally recruitment_pattern will expand into a recruitment_instructions with pattern=yes and importance=1. (So recruitment_pattern is a shortcut for a special recruitment_instructions). Make sure that there is no recruitment_instructions with importance=1 when using recruitment_pattern.
• villages_per_scout=4 (int) A number 0 or higher which determines how many scouts the AI recruits. If 0, the AI doesn't recruit scouts to capture villages.

Making recruitment strong

The default aspects are chosen so it is fun to play against the AI. To let the AI recruit stronger this settings are recommended:

• recruitment_diversity=0.8 Let the AI recruit the best units only.
• recruitment_randomness=0 Let the recruitment not be random.
• villages_per_scout=0 Test shown that the AI plays slightly better with less or no scouts at all.
• deactivate recruitment_save_gold (see above). Test shown that the AI plays better when it spends all money immediately.

In Multiplayer games the AI 'Strong AI (RCA)' can be chosen where the settings are as described here.

Some notes on multiple leaders

The CA can recruit with multiple leaders - even with different recruitment lists. Also the CA 'Move_Leader_To_Keep' will try to move all units with canrecruit=yes to the next keep.

However there are some drawbacks:

• In the current version the leaders will recruit by default almost the same amount of units. This can partly configured with the aspect recruitment_instructions and it's attribute leader_id. A leader will recruit more units when he/she is in danger (enemies are near).
• As mentioned above leader_id in recruitment_instructions is only a recommendation.
• Multiple leaders with different recruitment lists in combination with a recruitment_pattern may be confusing. (For the AI it's more important that the leaders recruit the same amount of units than fulfilling the ratios given in the recruitment_pattern.)

See also: Wesnoth AI