Design of Emergent and Adaptive Virtual Players in a War RTS Game - Foundations on Natural and Artificial Computation - page 376

agents/soldiers). An army controller can be viewed as a set of rules that will
control the reactions of its agents according to the current state of play. De-
pending on its situation, the agent executes a particular action that may modify
the state of the agent itself. The global team strategy emerges as the result of
the sum of the specific behavior of all the agents.
However, the definition of specific strategies for each agent is complex, requires
a profound knowledge of the required behavior of each agent, and results in
predictable behavior. Moreover, devising independent controllers for each unit
in a RTS game is costly; in fact this is not realistic as requires a high computation
effort that surely decreases the quality of the real-time rendering of the game.
One solution consists of designing global team strategies; however, again this is
very complex as the programmer has to cope with too many possibilities that
arise from the different interactions among all the agents of both armies. We have
opted by a more realistic process: all the soldiers will be managed by the same
controller, and in any instant of the game, the team behavior will be the result
of summing all the action taken by each of its constituent units. This means that
we have to produce just one controller to devise a global team strategy. Note
however that this does not mean all the agents execute the same action because
the action to be executed by a soldier will depend on its particular situation in
the match.

3.2 A Two Phases Process

The procedure described in the following has the aim of creating a virtual player
whose behavior auto-evolves according to the skills that human player exhibits
along the game. Algorithm 1 displays the general schema of the process: initially
(in the first game), the virtual player (VP) is constructed as an expert system
(i.e., a rule-based prototype (RBP)) that comprises tens of rules trying to cover
all the possible situations in which an agent might be involved. This system was
the result of a number of trials designed from the experiences of the authors
playing wRTS games. Then, let

henceforth, and assume the
human player will play a number of ℘ games, the procedure consists of two phases
that are sequentially executed. Firstly, a player modeling phase is conducted;
this step is described in Section 3.3 and basically consists of building a model
of the behavior that human player exhibits during the game (i.e., on-line). The
second phase, described in Section 3.4, is devoted to construct the virtual player
by evolving a strategy controlling the functioning of a virtual army unit via
an evolutionary algorithm (EA). In the following we describe in details both
phases. Firstly, we discuss how the virtual player is internally encoded as this is
important to understand the functioning of the whole schema.

N h be { 1 ,...,h }

Representation: the virtual player contains a NPC army strategy, and its
internal encoding corresponds with both the rule-based expert system and any
individual in the EA population. In the following we will refer to 'individual'
in a general sense. Every individual represents a team strategy, that is to say,
a common set of actions that each army unit has to execute under its specific