Target Selection in RTS Battles

1
Target Selection in RTS Battles

• Doug Demyen

2
Outline

• Introduction
• Problem Setup
• Challenges
• Helpful Observations
• Static Strategies
• Fixed Priority
• Randomized

• Dynamic Strategies
• Genetic Algorithms
• Stochastic Hillclimbing
• Simulated Annealing
• Further Work
• Conclusion

3
Introduction

• Outcomes of battles in RTS games can vary greatly
  simply by selection of targets
• User wants a good outcome but might have other
  priorities than directing the battle
• Similarly, we want enemy AI to seem as if someone
  is directing the battles

4
Problem Description

• A battle is made up of a number of units
  belonging to one of at least two opposing armies
• Each unit has three attributes we are concerned
  with
• Hit points
• Attack Strength
• Cooldown Period

5
Problem Description (Contd)

• At each moment in the battle, each active unit
  that has not attacked for its cooldown period
  deals it attack strength of damage to a target
  unit
• That units hit points are reduced by this
  damage, and if they fall to or below zero, that
  unit is no longer active
• The battle continues in this way until only one
  army has any remaining units

6
Solution Description

• The solution for one army will involve the target
  for each of its units given any situation in the
  battle
• This is huge these possibilities cannot be
  enumerated
• Another problem is that in an RTS, we must
  determine targets quickly good targets mean
  nothing if we are dead!

7
Solution Description (Contd)

• Also the battle is seldom static rarely will
  all units join the battle at the same time and
  battle until the end
• Units can join a battle in progress, and can
  leave a battle before it is finished or they are
  eliminated
• Unexpected events could change the battle like a
  bomb dropped on some units
• Want an approach that can handle this

8
Helpful Observations

• Generally we want persistent targets
• Once we target a unit, we should target it until
  it is eliminated
• This is because eliminating units stops them from
  inflicting damage on allied units thus a more
  fit solution
• Obvious departure from this is when the targeted
  unit will be destroyed this turn anyway

9
Helpful Observations (Contd)

• Also, events many time steps ahead in the battle
  seldom affect the desired solution
• We still want to maximize the damage and
  casualties inflicted versus that sustained,
  regardless of the time
• For example, doing poorly for some part of the
  battle does not help to win the overall battle

10
A More Compact Solution

• This allows us to represent an effective solution
  as follows
• For each allied unit, define an attack queue
  where it targets each unit in the queue until it
  is destroyed, then moves on to the next
• The queues will be used for a certain time
  (several seconds of the battle), so they only
  need to hold a few enemy units

11
Attack Queue Example
2 3 1
4 1 3
2 4 1
4 2 3
Attack Queues
2
3
4
1
Indices
12
What is the Goal?

• Should the goal be eliminating the opposing army?
• If 97 allied units destroy 3 enemy units with 50
  of them firing wildly into the air, is that
  success?
• If those 3 units destroy 20 units in the opposing
  army before they are destroyed, did they fail?

13
A Definition of Fitness

• The goodness of a target designation or its
  fitness will be defined as
• The amount of damage totally dealt to opposing
  units plus a bonus for each enemy eliminated
• Minus the total amount of damage sustained by
  allied units and an additional penalty for
  allied units lost

14
Approaches

• Some static strategy for example, aiming every
  unit at the active opposing unit that poses the
  greatest threat
• A more dynamic approach evaluate different
  target designations and try to find the best one

15
Fixed Strategy

• Define an opposing units threat to be the ratio
  of its damage per turn (attack strength /
  cooldown period) to hit points remaining
• The opposing unit with the highest such ratio is
  targeted by all units
• If the units already targeting this unit is
  sufficient to destroy it, aim subsequent units at
  the enemy with the next highest threat, and so on

16
Randomized Strategy

• Using the threat definition in the last slide,
  the randomized version
• Aims each allied unit randomly at one of the
  opposing units with the highest assessed threat
• Cant determine when to pick a different unit to
  target because others will destroy it this turn
• Used for comparison purposes

17
Evolutionary Algorithms

• This type of solution lends itself to
  evolutionary algorithms
• Have the advantage of working on a solution and
  returning the current best whenever its out of
  time
• Since solutions are determined every few seconds,
  they can also be determined if the battle changes

18
Genetic Algorithm

• For n allies and attack queues of length k,
  encode it as a genome of length n k
• Random genes for mutations are indices of
  opposing units that are active and do not appear
  earlier in the same units attack queue
• Crossovers have their normal meaning

19
Stochastic Hillclimbing

• With a target list encoded similarly to a genetic
  algorithm genome
• Generate a neighbour of it by changing a target
  (similar to mutating)
• Test the fitness of this neighbour
• If it is more fit, replace the current solution
  by this new one
• Iterate until time has expired

20
Simulated Annealing

• Similar to stochastic hillclimbing, except
  solutions that are not as fit can be accepted
• Probability of this happening decreases with more
  iterations
• Less likely to get stuck in a local maximum,
  but might produce a poor solution if not allowed
  to converge

21
Further Work

• Extend to ORTS-like system of different kinds of
  attack (piercing, explosive, etc.) and different
  defenses
• Incorporate attacks with non-constant damage
• Take into account inter-unit distances, attack
  ranges, and unit movement

22
Conclusion

• In light of the challenges of target selection in
  RTS battles, I feel evolutionary algorithms
  provide
• Flexibility for a changing environment
• Ability to produce a good solution quickly
• Improved solutions given more time
• More comparisons still to come . . .

23
References

• Alan Watt and Mark Watt. AI Techniques for Game
  Programming. Premier Press, 2002.
• Alexander Kovarsky. Heuristic search applied to
  abstract combat scenarios. Masters Thesis,
  University of Alberta, 2004.
• M. Tim Jones. AI Application Programming. Charles
  River Media, 2003.

24
Questions?

• Thank You