Artificial Agents
27
Reinforcement learning can be used to teach agents correct behavior. A
simulation game ‘Black and White’ [191], used reinforcement learning to
teach characters the difference between evil and good. The game allowed
players to act like gods, with the capability of controlling creatures.
Alternately, evolutionary programs in games can also be taught to allow
new state machines to modify or evolve to generate more intelligent
programs (e.g. in a game Rougelike [50]). To achieve this, a reward
structure is used in conjunction with evolutionary algorithms to modify
behavior, with reward acting as a payoff (fitness).
However, reinforcement learning, in itself, is very limited as it focuses
only on agent performance. Agents can use it to choose different behaviors and modify a set of allowable actions to optimize behavior, adapting
at t.
Other research in evolutionary concepts in computer science are summarized below,
Turing [195]. Recognized the connection between evolution and machine learning.
Friedman [69]. Proved thinking machine can be used in playing chess
games.
Friedberg [68]. Improved search space for good programs with given
possible solutions.
Bremermann [30]. Presented a multi-objective solution to a numerous parameters in a function. “to a stable point ... [which] need not
even be a saddle point [of the fitness function].”
Reed [158]. Used evolutionary algorithms in poker games. Presented
use of crossover to find quicker solutions.
Minsky [133]. Objected to Friedberg’s solutions saying that they take
too much of time to compute.
Fogel [62] [63]. Combined finite state machine with payoff function for
producing evolving machines.
Fogel and Burgin [66]. Introduced evolutionary concepts for gaming.
Rechenberg and Schwefel [156] [173]. Produced
evolutionary
strategies.
Holland [88] [89]. Worked on genetic algorithms for adaptive system.
B¨ ack and Schwefel [15]. Compared results of experiments for varying crossover and mutation rates.
Turing [195] showed how evolution can aid machine learning by generating new state machines through trial and error. While, Friedman [69]
coined term ‘thinking machines’, using mutation and selection methods
in evolutionary processes to give birth to new machines. These efforts
27
Reinforcement learning can be used to teach agents correct behavior. A
simulation game ‘Black and White’ [191], used reinforcement learning to
teach characters the difference between evil and good. The game allowed
players to act like gods, with the capability of controlling creatures.
Alternately, evolutionary programs in games can also be taught to allow
new state machines to modify or evolve to generate more intelligent
programs (e.g. in a game Rougelike [50]). To achieve this, a reward
structure is used in conjunction with evolutionary algorithms to modify
behavior, with reward acting as a payoff (fitness).
However, reinforcement learning, in itself, is very limited as it focuses
only on agent performance. Agents can use it to choose different behaviors and modify a set of allowable actions to optimize behavior, adapting
at t.
Other research in evolutionary concepts in computer science are summarized below,
Turing [195]. Recognized the connection between evolution and machine learning.
Friedman [69]. Proved thinking machine can be used in playing chess
games.
Friedberg [68]. Improved search space for good programs with given
possible solutions.
Bremermann [30]. Presented a multi-objective solution to a numerous parameters in a function. “to a stable point ... [which] need not
even be a saddle point [of the fitness function].”
Reed [158]. Used evolutionary algorithms in poker games. Presented
use of crossover to find quicker solutions.
Minsky [133]. Objected to Friedberg’s solutions saying that they take
too much of time to compute.
Fogel [62] [63]. Combined finite state machine with payoff function for
producing evolving machines.
Fogel and Burgin [66]. Introduced evolutionary concepts for gaming.
Rechenberg and Schwefel [156] [173]. Produced
evolutionary
strategies.
Holland [88] [89]. Worked on genetic algorithms for adaptive system.
B¨ ack and Schwefel [15]. Compared results of experiments for varying crossover and mutation rates.
Turing [195] showed how evolution can aid machine learning by generating new state machines through trial and error. While, Friedman [69]
coined term ‘thinking machines’, using mutation and selection methods
in evolutionary processes to give birth to new machines. These efforts