Information Technology Reference
In-Depth Information
Kok et al. [51, 52, 53] proposed the use of a Coordination Graph (CG) as team-
work model for agents to select an optimal joint action and coordinate the execution
of their individual actions. This approach requires that each agent communicated his
local payoffs for executing a set of individual actions with neighboring agents in order
to find a joint action that maximizes the payoff for the team. An extension of this
work was proposed to render communication superfluous [51, 53] by assuming that i)
players can identify one another; ii) payoff functions are known by all teammates; iii)
players can compute the adequacy to fulfill a role for all others; iv) the order of ac-
tions is known among all players; and v) in context-specific CGs, all players reacha-
ble from a given player can observe the state variables located in his value rules.
Other methods for coordination made use of an explicit communication of belief
states [54, 55, 56]. These methods differ on the following criteria: i) communication
is assumed to be flawless in [54] but not in [55]; and ii) utility measure of other agents
states is estimated based on a sum of heuristic functions represented as potential fields
[54] and in [55] it is based on meaningful observed experience.
Stulp et al. [56, 57] proposed the use of temporal prediction models of teammates
to enable agents to coordinate for regaining the ball possession. These models are
learned offline, using model trees and a neural network, based on the observation of a
match with no opponents. These are used to estimate the state of teammates and antic-
ipate the utilities of their intentions in order to adapt to their predicted actions.
Desheng and Kejian [58] proposed a non-communicative approach for coordinat-
ing a team of agents using the notions of roles and situation calculus. A role consists
on a description of a teammate task and is regarded as a form of intentional coopera-
tion. This approach assumes that an agent is able to compute the adequacy of each
agent to fulfill all possible roles based on the situation calculus and consequently
figure out the roles that all agents will adopt. Using common knowledge about roles,
each individual agent is able to predict the actions of others based on the situation
calculus and render communication superfluous.
3.2
Communication Languages
The definition of coaching languages was driven by the need to convey advices from
coaches to players during a match and to be able to interact with heterogeneously
designed players. The languages Coach Unilang [11], CLang [59], Strategy Formali-
zation Language (SFL) [60] were proposed to structure this communication.
Coach Unilang was proposed on 2001 by Reis and Lau [11] as a generic coaching
language enabling high-level coaching of a (robo) soccer team. The language in-
cluded all features that enable to coach a robo soccer team such as tactics, formations,
actions, player types, conditions, regions, periods among others.
CLang [59] was based on the initially proposed Coach Unilang [11] as a simplifi-
cation in order to be the standard language used by coaches in the RoboCup 2D
simulated league and enable to promote a competition focused on simple low-level
coaching techniques. It latter evolved to integrate most of the features of Coach Un-
ilang and be a generic coaching language. On this latter version, tactics and behaviors
are described using rules which map
directives
(lists of
actions
to execute or avoid) to
conditions
(match situations descriptions). A
condition
is a logical expression based
on game variables (e.g. objects positions) whereas an
action
is a low-level skill
