Difference between revisions of "Adventure AI"

From VCMI Project Wiki
Jump to: navigation, search
(Goal decomposition)
(Goal decomposition)
Line 7: Line 7:
  
 
= Goal decomposition =
 
= Goal decomposition =
AT eahc step of algorithm, AI chooses one best sub-goal that may lead to completion of current goal (function striveToGoal). Decomposition is performed until there is possibility to complete the goal directly (goal is elementar). Typical elementar action is moving hero to certain tile.
+
AT each step of algorithm, AI chooses one best sub-goal that may lead to completion of current goal (function striveToGoal). Decomposition is performed until there is possibility to complete the goal directly (goal is elementar). Typical elementar action is moving hero to certain tile.
 
Currently the first matching goal is chosen and set as target. Then AI tries to accomplish it. End of action is indicated by exception. After making an action, AI continues to decompose the main goal until all possibilities are exhaused (infinite while loop).
 
Currently the first matching goal is chosen and set as target. Then AI tries to accomplish it. End of action is indicated by exception. After making an action, AI continues to decompose the main goal until all possibilities are exhaused (infinite while loop).
 
Some goals can be abstract. When abstract goal is chosen, it becomes main goal itself. This way abstract goal can result in multiple actions from AI, until all possible actions are exhausted.
 
Some goals can be abstract. When abstract goal is chosen, it becomes main goal itself. This way abstract goal can result in multiple actions from AI, until all possible actions are exhausted.

Revision as of 11:38, 29 December 2013

VCAI is the module responsible for player actions on adventure map.

Basic activity

AI starts with "WIN" ultimate goal and tries to accomplish it, based on win condition for certain scenario. Additionally, it performs typical actions (like building in town) each turn. THis should be replaced with actual reasoning. At the end of each turn, heroes who have remaining movement points try to visit all unvisited objects within their range (wander function).

Strategy

Additional goals can be saved to be completed in the future. There is a list of locked heroes who are tied to particular important goals and try to realize them each turn. There is a list of reserved objects tied to heroes. If a hero tries to visit an object but can't reach it this turn, object is reserved. Other heroes will not try to visit that object until it's reached (or reserving hero is lost).

Goal decomposition

AT each step of algorithm, AI chooses one best sub-goal that may lead to completion of current goal (function striveToGoal). Decomposition is performed until there is possibility to complete the goal directly (goal is elementar). Typical elementar action is moving hero to certain tile. Currently the first matching goal is chosen and set as target. Then AI tries to accomplish it. End of action is indicated by exception. After making an action, AI continues to decompose the main goal until all possibilities are exhaused (infinite while loop). Some goals can be abstract. When abstract goal is chosen, it becomes main goal itself. This way abstract goal can result in multiple actions from AI, until all possible actions are exhausted.

Hero goals

If returned goal is related to hero, this goal is then assigned to specific hero. Heroes continue their goals in subsequent turns, until it's complete. Every action that may result in completing teh goal must inform AI about such opportunity so that hero goal is erased.

List of goal types

Fuzzy evaluation

Since 0.94c AI used fuzzy logic to compare the multiple goals. Each goal returns the list of all possible subgoals via getAllPossibleSubgoals function. Then these goals are evaluated and best sub-goal is chosen for further processing. Each goal holds its last calculated value in "priority" field.

Locked heroes and preemption

Priorities of locked goals for heroes are stored and used in fuzzy comparison. The greater the priority, the less likely it is to choose particular locked hero. However, preemption is always possible. This way no heroes would get stuck on their goals if they are found outdated and suboptimal in new situation.