Heuristic Search in Empire-Based Games

1
Heuristic Search in Empire-Based Games

• CS-470
• Nathan Freeburg

2
Introduction

• Goal
• Develop a hierarchical heuristic search algorithm
  to determine moves for a simple strategy game.
• Why heuristic search?
• Why hierarchical?

3
Heuristic Search

• What is it?
• Method of choosing branches in a state space
  which are more likely to lead to an acceptable
  solution to the problem.
• Often implemented through minimax search.

4
Search Example
A
A
2
2
D
C
B
D
C
B
2
-99
-2
2
-99
-2
M
L
K
J
I
H
G
F
E
M
L
K
J
I
H
G
F
E
2
4
5
-99
-1
1
-2
2
4
2
4
5
-99
-1
1
-2
2
4
5
Reducing the Search Space

• A full search for a game with n pieces and m
  moves per piece, to a depth of d, requires
  examination of (mn)d separate states.
• A full depth search for each piece, performed one
  at a time for all pieces, reduces the number of
  states to be examined to n(m(d1)).
• Remember moves made by other pieces at the
  current depth.

6
Algorithm Used

• Choose a piece to focus on.
• For all other pieces, perform a depth 1
  lookahead to determine best move at that level.
• Perform a full depth lookahead for the chosen
  piece to determine the best move.
• Repeat for all pieces.

7
Constructing a Hierarchy

• Split the decision making process into levels.
• Determine overall strategy at the highest level.
• Rule-based approach.
• Determine the moves necessary to implement that
  strategy at the next level.
• Heuristic search.

8
Test Game

• Loosely based on Empire, via Empire Lite.
• Only resources are cities.
• Advantage goes to attacker.
• Two win conditions.
• All opposing units eliminated.
• All cities occupied by one side.

9
Results

• Hierarchical heuristic search outperformed
  non-hierarchical search.
• Average move time was about 5 seconds for a
  lookahead depth of 8 on a 400 Mhz processor.
• Still too high for mainstream applications, but
  promising.

10
Further Work

• At a superficial level further tuning of the
  heuristics and tightening of the code.
• Division of units into separate groups, each of
  which could be given different goals.
• More robust set of rules for decision making at
  the top level.

11
Conclusions

• Hierarchical heuristic search preferable to
  non-hierarchical, despite increase in complexity.
• Use in mainstream applications dependent on
  further increases in processing speed.