Automated Terrain Analysis in Real-Time Strategy Games

Full text

Automated Terrain Analysis in Real-Time
Strategy Games
Chen Si
Games Studio
University of Technology, Sydney
chen.si@student.uts.edu.au
Yusuf Pisan
Games Studio
University of Technology, Sydney
yusuf.pisan@gamesstudio.org
Chek Tien Tan
Games Studio
University of Technology, Sydney
chek@gamesstudio.org
ABSTRACT
Real-time strategy (RTS) games represent a mainstream genre of
video games. They are also practical test-beds for intelligent
agents, which have received considerable interest from Artificial
Intelligence (AI) researchers, in particular game AI researchers.
Terrain knowledge understanding is a fundamental issue for RTS
agents and map decomposition methods can help AI agents in
representing terrain knowledge. These contributions support AI
agents’ path finding and combat strategy. In some RTS games,
such as StarCraft, all terrain information is provided to AI agents
at the beginning of the game. This presents an unfair advantage,
as human players do not have access to this information. We
propose a terrain analysis framework, in which AI agents gather
terrain knowledge by managing scouts to explore game maps.
This framework is part of my Ph.D. study that is investigating
scouting strategies for RTS games. We developed an extension to
the StarCraft system, called terrain engine that releases terrain
information in small chunks rather than providing the full map, to
investigate human-like techniques for scouting. Within the terrain
analysis framework, we present a reconnaissance (recon)
algorithm to guide individual scout units in recon tasks. Then, we
identify the factors for terrain exploration planning model, which
will be implemented as part of our future work.
1. INTRODUCTION
Real time strategy (RTS) games pose interesting challenges for
both human players and artificial intelligence (AI) bots. In RTS
games, a player need to consider high-level game issues such as
building order, unit composition and troop formations, and at the
same time micro-manage individual units to attack efficiently in
packs and to minimize damage. Unlike traditional turn-based
board games such as Chess and Go, in which players are able to
get access to full game-state, RTS games require players to make
decisions with partial game-state information. Actively gathering
information about the opponent is a crucial strategic component
for RTS games.
Terrain information collection and analysis is an important
foundational part of RTS games, due to the vital information it
contains. The analysis results [12] are used in strategy making for
AI bots (e.g. ambushing in narrow paths) [3] and in advanced
path-finding algorithms (e.g. navigating unit groups) [13].
A lot of research has been performed in terrain information
analysis. AI agents can recognize general terrain features such as
regions, chokepoints and base-locations by applying computer
vision techniques to game maps [3][11]. These techniques rely on
having access to the full game-map at the beginning of the game,
and would be considered cheating for games where human players
do not have access to this data. Although strong players know the
popular maps, most novice players are unfamiliar with the game
maps. A cheatings AI [8], which unfairly gets access to global
game information, destroys the game experience for players.
As part of my Ph.D. study, we present a dynamic terrain analysis
framework that collects map information and recognizes terrain
features by managing scout units tactically and strategically.
StarCraft Brood War is a suitable platform to test our idea about
scouting in unknown territory. We, thus, describe a terrain engine
that has been implemented as an extension of StarCraft Brood
War. The terrain engine releases terrain information based on
what AI agents scouting units would be allowed to detect in their
immediate vicinity. We present a novel tactical navigation strategy
for scouting. As a part of our future work, we are also exploring a
planning module to determine when to scout, where to scout and
which units are chosen to scout, by considering a variety of
factors such as size, movement speed and price of units. Our
novel terrain analysis method reasonably avoids cheatings, and
contributes to build up human-like AI robot in RTS games.
Scouting in unknown territory presents a lot of challenges, such as
balancing resource-consuming between exploring terrain and
attacking enemy. In this paper, we are investigating how to
explore terrain features. Overall scouting strategies for exploration
are left for our future work.
2. RELATED WORK
Terrain understanding is a fundamental task for RTS AI agents to
extract essential information for other decision-making sub-
systems. Efficient analysis methods have been developed to
represent game maps in different ways. A title-based terrain
representation system has been used in BANG – an RTS game
[12]. Game maps are divided into tiles and classified into two
types – convex areas and non-convex areas, which are used in
tracking routes. According to the requirements of combat
scenarios, game space [3][11] was divided into regions (free
movement areas), chokepoints (narrow areas) and obstacles, by
using image processing techniques and the Voronoi diagram. The
Hale et al. [5] also presented an automatic growth mesh technique
to perform region partition – DEACCON, in which small squares
are seeded and grown to detect edges. This automated navigation-
mesh-generation method has also been extended into 3D scenarios
[6].

The works mentioned above focused on the analysis of pre-loaded
game maps. In this paper, we take a different approach: building
map information incrementally. Navigation of scout units is the
primary challenge for our terrain exploration agent. Potential field
technique [4] has been used to deal with fog of war in Wargus (a
clone of WarCraft 2), evaluating unexplored terrain tiles to
navigate scouts to explore terrain. Park et al. [10] presented a
heuristic navigation tactic for scout units in collecting opponents’
information. They devised a navigation method where a scout
walks around the enemy base.
In terms of planning in uncertain game states, several tactical
planning algorithms provide references for us to develop a
strategic terrain-exploration planner. For example, Chung et al. [2]
employed the Monte Carlo method to create a random planning,
evaluation and evolving planner for combat scenarios. Another
promising algorithm [1] – UCT (a kind of Monte Carlo planning)
was used in making tactical assault plans, concentrating on group-
based plan making. On the resource collection aspect, Naves and
Lepes [9] presented a stochastic search and planning method to
solve the resource production-planning problem. Many strategies
for RTS games play an important role. All of these strategies now
still rely on full map information being available. Being forced to
do scouting as well as choosing and modifying a strategy makes it
more challenging.
3. METHOD
Our proposed AI agent framework consists of four main parts as
shown in Figure 1. In order to plan the recon task, the Exploration
Planning module first performs Scout Selection, using known
terrain data and other relevant game information. Then it assigns
the scouts to appropriate destinations and informs the Navigation
component to conduct low-level guidance. Next, the Dynamic
Terrain Database module is an extended version of the
Knowledge Database, which adds various other types of
exploration-guidance data. The Patch Management component is
utilized to manage explored patches (small map chunks), and to
recognize decomposed terrain information from these patches.
The Navigation component then leads the way of a scout unit to
explore a certain area efficiently.
In many RTS games such as StarCraft: Brood War, the game
system does not provide the full maps. Therefore, analyzing the
full map, which is not normally available to human players,
provides an unfair advantage to AI agents. We have hence
implemented a Terrain Engine that releases limited information to
AI agents cumulatively, rather than releasing the full map at once.
The Terrain Engine provides a platform for our framework to
explore scouting strategies for AI agents in a fair setting, similar
to how human players employ scouting. The Terrain Engine will
only release information about parts of terrain that are in the
visible area of a scout. There are two kinds of information in the
Knowledge Database of the Terrain Engine. One is a regional
database, which maintains the results of terrain decomposition,
including regions, chokepoints, and locations of resources. The
other is a tile-based database, in which a game map is represented
by square tiles. Each tile is labeled passable or impassable,
depending on whether a ground unit can pass or not.
The game system releases terrain information to AI agents based
on the four transmission rules below:
(1) The engine will release information, including passability of
tiles, points in edges of regions (edge-points), end-points of
chokepoints and center-points of resources, when they are in
the visible area of a scout.
(2) If all the edge-points of a region have been detected by
scouts, information of the entire region will be released.
(3) If two end-points of a chokepoint have been explored,
information of the chokepoint will be released.
(4) If center-points of resources are “seen” by scouts, we
assume that they know the entire information about the
resources (e.g. locations and reserves).
3.1 Exploration Planning Module
The purpose of the Exploration Planning module is to collect the
most valuable terrain data for the current game state as soon as
possible with minimal resources. Therefore, deciding which
locations to explore and choosing appropriate scouting units in
corresponding game states are significant challenges for RTS
agents.
Some areas are more valuable than others during certain game
periods. For example, areas around the straight-line path from a
player’s start location to the enemy base are frequently visited
(hence valuable) by units during beginning 10-15 minutes.
Therefore, scouts should preferably consider exploring these areas
first.
The role of the Exploration Controller is to generate the
exploration plans. In an exploration plan, the general procedure of
exploration starts by assigning a scout to explore the possible
start-locations of the enemy base, right when the game starts.
After that, the scout walks along its movement direction, until one
edge of the region is “seen”. Then, the Navigation component
(Section 3.5) takes over control of the scout in order to explore
the region. Next, the following steps are repeated: (1) Search a
valuable area for current game state, based on obtained terrain
information. (2) Manage a scout to move to one edge of the area,
and active the domination of the Navigation component.
The role of the Scout Selection component is to consider which
units should be assigned for recon tasks, in which several
properties of the scout units need to be considered.
(1) Visibility: Different units have different sight ranges. Units
with larger sight ranges are able to “see” further.
(2) Speed: Units with higher movement speeds can perform
scouting faster.
Figure 1. An overview of terrain analysis.
.
Knowledge Database
Transmission Rules
Terrain Engine of Game System
Scout
Selection
Exploration
Controller
Exploration Planning
Terrain
Decomposition
Information
Terrain
Tiles
Information
Dynamic Terrain Database
Navigation
Scout
Units
Patch Management
AI agent
Game Platform
Control Flow
Data Flow
Legend

(3) Armor: Stronger armor can help scout units live longer in
dangerous scouting areas.
(4) Cost: The cost for producing scout units varies. The cost
should be balanced between producing more un-expensive
scouts to do multiple units scouting and using less advanced
scouts to complete the tasks.
(5) Air or ground: Air units, when available, can be used for
scouting when the situation requires.
3.2 Data Structure
We combine two kinds of data structures, region-based structure
and tile-based structure, to represent RTS game terrain for
exploration planning and navigation purposes separately. Region-
based structure classifies areas into three different types: regions,
chokepoints and obstacles. In this representation method, we
employ region, chokepoint, baselocation and startlocation to
record essential elements of a game map, like BWTA [11] did. A
region structure contains the data of its border polygon and
indexes of connected chokepoints. In a chokepoint structure, two
endpoints of the chokepoint and the indexes of the two connected
regions are presented. A baselocation is a kind of region with
additional information about resources contained in that region. A
startlocation is a baselocation with a player’s information.
This structure is an efficient way for the Exploration Planning
module to infer the next valuable area to be explored. For
example, when the startlocation of the opponent is found, areas,
which can be entered via chokepoints from the startlocation, are
more likely to be explored, if resources exist in these areas.
Furthermore, data contained in this structure also plays an
important role in making strategies. For example, a player would
tend to choose a rush strategy if the distance between
startlocations are short, and chokepoints around them are
relatively wide.
The tile-based structure divides a game map into tiles with the
size of the smallest unit. Based on obstacle information, a tile is
also flagged with passable or impassable. In this paper, our
navigation algorithm employs the tile-based information to guide
a scout in exploring an area. Data in tiles also contributes to more
efficient path finding.
3.3 Navigation Component
The purpose of the Navigation component is to guide a scout unit
to collect data in an area. Generally, the navigator takes control of
a scout unit, and begins to navigate it when the unit arrives at the
designated spot. The navigator follows two principles. (1) Scout
units need to avoid damage from enemy units and keep alive. The
resources for scouting are extremely precious, since AI bots may
weaken their economy, if they assign too many units to scout.
Furthermore, scout can potentially collect more data if they can
stay alive longer in unexplored areas. (2) Units should explore
designated areas as soon as possible.
To adhere to these two principles, a scout unit explores an area by
mimicking human to scout for practical purpose in our system. It
will walk around a region whilst keeping an optimal distance from
the edge. We employ the tile-based structure to design the
navigation algorithm. A tile can be one of three types ‒ passable,
impassable and unknown. Initially, all tiles are flagged unknown.
When they are gradually uncovered, their types change to
passable or impassable, and this change only happens once.
We use a potential field value recorded in each tile to help units
find an optimal path in exploring. The method is based on [4][7].
The value is calculated when passability of tiles has been explored
in the visibility area of a scout for each frame. The algorithm for
calculating this value is shown below:
Algorithm 1 Calculate Tile Potential Value
Require : t ∈ gridmap and sr ∈ S
square ← t
pv ← 0
if passable(square) then
while passable (square) do
pv ← pv + 1
square ← expand (square)
if pv > (sr - 2) then
pv ← 0
break
else
pv ←-1
return pv
t refers to a tile in the map. pv represents the potential value of tile
t. sr is the sight range of current scout unit. passable() is used to
check whether a square is passable or not. expand() is to expand a
square into its adjacent tiles in eight directions. The higher the
potential value of a tile is, the more possible that the scout will
move to the location of the tile in next step.
To prevent scout units visiting tiles repeatedly, we include a
visited flag to record tiles that have been reached in the last
several steps. Visited tiles are excluded when choosing the tile
with the highest potential value for the next destination of a scout
(among the eight adjacent tiles). Sometimes, a scout may be
confused when there are two or more tiles with the same highest
potential value. In this situation, the previous directions of
movement will be used as tiebreakers. Vectors from the current
tile to these candidate tiles are first generated. The eventual
selected destination tile is then the candidate tile with the smallest
angular difference between its vector and the previous direction
vector.
3.4 Patch Management Component
The terrain data that is captured by a single unit in each frame will
be known as a patch in our system, since the visible area of a
single unit can only cover parts of a region in most cases. As data
is often incomplete, per-frame queries for recognizing terrain
features might be a waste of time. Moreover, terrain data in
sequential frames almost always overlap with each other. These
two vital problems, (1) how to organize patches properly without
overlaps and (2) how to query terrain features efficiently, are
handled in our Patch Management component.
In our system, we develop a hierarchical Patch Management
component. In each frame, points of terrain data that a unit has
collected in its visible area are organized into patches, according
to their spatial connectivity between each other. The patches are
stored in a local patch buffer. Patches are merged, if they are
spatially connected, when they are pushed into the local patch
buffer. The patches in the temporal patch buffer are transmitted
and merged into a global patch buffer, in which global patches are
stored, for a shot period. After merging, data in the local patch
buffer is cleared. This method hence helps to reduce redundant

data, since overlapped patch data is merged and cleared when
necessary, and to improve efficiency, because the system do not
need to traverse the global patch buffer to merge the new patch in
each frame. For each long period, terrain features are recognized
by comparing patch data in the dynamic patch buffer and region
data in the Terrain Region Database. Recognized terrain features
are finally stored in the Terrain Region Database, and
corresponding patches are deleted in the dynamic patch buffer.
4. IMPLEMENTATION
Our framework is implemented in StarCraft: Brood War, one of
the most popular RTS game genres in the last decade. It is coded
in C++, based on the Brood War API (BWAPI)
1
– a library to
provide interfaces for game AI development. The terrain game
engine uses the Brood War Terrain Analyser (BWTA) [11] to
analyse game maps, and thus obtain all of its processed terrain
information. This system can be easily embedded into other AI
agents, which are able to play full games. We have finished
designing and implementing the Terrain Engine and Navigation
component. The implementation of Patch Management
component and Dynamic Terrain Database is expected to be
completed soon. When these low-level and foundational
components are completed, we will design the exploration
planning algorithm and embed it into the Exploration Planning
module.
5. EXPERIMENT
Our framework will be evaluated by comparing its performance
with the BWTA. We will first choose three AI bots from previous
AIIDE StarCraft competitions
2
, which rely on BWTA to obtain
terrain data. These bots will then be modified into three different
versions: (1) the original version, (2) a version that knows nothing
about the game terrain, by cutting its link with BWTA, and (3) the
version that contains our scouting framework. For each bot, its
three versions will compete against each other with 15 games in
each map. Nine ICCup (International Cyber Cup) maps will be
chosen for this experiment. We will then compare the win-rate to
identify the practical applicability of our detection model in RTS
game agents. Moreover, we will invite 20 novice players, who
know nothing about maps that we choose above, but they have
basic skills to play StarCraft: BroodWar. These human players
will be organized to play against version (1) bots and version (3)
bots five times separately in each map. The questionnaires will be
made to test whether our method helps to increase the game-
playing experience.
6. CONCLUSIONS
Terrain knowledge plays a pivotal role in helping RTS game
players win by using it to infer possible enemy locations,
determining opponent strategies and devising strategies that take
advantage of terrain features. Scouting is a believable way to
collect terrain data for AI players. In this paper, we demonstrate a
terrain analysis framework to effectively explore RTS game map,
and to distribute information interactively. Potential field method
and heuristic information are employed for navigation. The patch
management technique is used to manage explored terrain data
and to identify features from current data.
1
http://code.google.com/p/bwapi/
2
http://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp
Our current terrain analysis framework does not take into account
hostile units, whether they should be avoided or engaged.
Similarly, obstacles placed by the opponent that block exploration
have to be taken into account when scouting. The design and
implementation of the exploration planning algorithm are the
primary future work for this PhD.
7. ACKNOWLEDGMENTS
The authors would like to thank the financial support provided by
the China Scholarship Council, the University of Technology,
Sydney, and the Centre of Human Centred Technology Design.
8. REFERENCES
[1] Balla, R.-K. & Fern, A. 2009. UCT for Tactical Assault
Planning in Real-Time Strategy Games. In Proceedings of
the International Joint Conferences on Artificial Intelligence
(Pasadena, California, USA, July 11-17, 2009). IJCAI '09.
AAAI, 40-45.
[2] Chung, M., Buro, M. & Schaefer, J. 2005. Monte Carlo
planning in RTS games. CIG '05. IEEE, 117-124.
[3] Forbus, K. D., Mahoney, J. V. & Dill, K. 2002. How
qualitative spatial reasoning can improve strategy game AIs.
Intelligent Systems, 17, 4, 25-30.
[4] Hagelback, J. & Johansson, S. J. 2008. Dealing with fog of
war in a real time strategy game environment. CIG '08. IEEE,
55-62.
[5] Hale, D. H., Youngblood, G. M. & Dixit, P. N. 2008.
Automatically-generated convex region decomposition for
real-time spatial agent navigation in virtual worlds. AIIDE
'08. AAAI, 173-178.
[6] Hale, D. H. & Youngblood, G. M. 2010. Automated
navigation mesh generation using advanced growth-based
techniques. In Game programming gems 8, A. Lake, Ed.
Course Technology Press, USA, 244-255.
[7] Harabor, D., & Botea, A. (2008, December). Hierarchical
path planning for multi-size agents in heterogeneous
environments. CIG'08. IEEE, 258-265.
[8] Laird, J. & VanLent M. 2001. Human-level AI's killer
application: Interactive computer games. AI magazine 22, 2:
15.
[9] Naves, T. F. & Lopes, C. R. 2012. Maximization of the
resource production in RTS games through stochastic search
and planning. IEEE International Conference on Systems,
Man, and Cybernetics (Seoul, Korea, October 14-17, 2012).
SMC '12. IEEE, 2241-2246.
[10] Park, H., Lee, K., Cho, H.-C. & Kim, K.-J. 2012. Prediction
of early stage opponents strategy for StarCraft AI using
scouting and machine learning. In Proceedings of the
Workshop at SIGGRAPH Asia (Singapore, Nov. 28 – Dec. 1,
2012 ). WASA '12. ACM, 7-12.
[11] Perkins, L. 2010. Terrain analysis in real-time strategy
games: an integrated approach to choke point detection and
region decomposition. AIIDE '10. AAAI, 168-173.
[12] Pottinger, D. C. 2000. Terrain analysis in realtime strategy
games. CGDC '00.
[13] Preuss, M., Beume, N., Danielsiek, H., Hein, T., Naujoks, B.,
Piatkowski, N., ... & Wessing, S. 2010. Towards intelligent
team composition and maneuvering in real-time strategy
games. Computational Intelligence and AI in Games, IEEE
Transactions on, 2, 2, 82-98.