Although artificial intelligence enthusiasts can mimic the activity of simple ensembles of interconnected nerves, they have been unable to imitate the kind of intelligence necessary to get along in the real world. Even the most elementary information processing done by organisms seems quite beyond the capacities of programmed automata. Real-world intelligence demands skills that include moving about, perceiving space and time, recognizing self and non-self, and making sense of fragmentary input. During everyday activity animals notice unusual. They are especially quick to sense danger - in a hazardous environment, time may bee too short for deliberate thought. An animal's intuitive knowledge provides almost instantaneous recognition of potential dangers and possibilities of reward, and an ability to anticipate events in an ever-changing world.
Machine intuition
Сообщений 1 страница 5 из 5
Поделиться22012-05-03 19:25:43
Chess play also takes place in a dynamic environment. Like an animal in the wild, the expert player is aware of what is taking place on the board, quickly notices opportunities to seize an advantage, and stays attuned to potential dangers.
World Champion Capablanca is said to have remarked: "I know at sight what a position contains. What could happen? What is doing to happen? You figure is out, I know it!" Although few players will claim this level of awareness, their chess play is guided nonetheless by what they perceive is likely to happen.
The uneasy engendered by a sense of, say, back-rank vulnerability, can easily save a game when time is too short for a good think, or when exact calculation is impossible.
Поделиться32012-05-03 19:27:10
The trick underlying chessplay - and life - is recognizing what is important. Distinguishing relevant from irrelevant input is the chief information - processing activity of an organism, for one's continuing existence depends on accurate judgement even when significance is disguised, as though protective coloration. But coaxing a machine to see importance has proved difficult, so difficult that one suspects that we must be doing something wrong. Perhaps the problem lies in the presumption that training can take place in isolation, detached from the environment in which the machine must function.
Поделиться42012-05-03 19:49:37
The Massachusetts computer scientist Michael Kuperstein is trying to train an artificial neural net to imitate a baby's learning. His patented robot INFANT, with video cameras to determine limb position and to locate the objects it manipulates, is set loose in a real-world-like environment to train itself. Like its organic role-model, the robot learns by exploring, by experiencing the consequences of its actions, and by repeating favorable results. It follows no preprogrammed coordinates, but conforms instead to it's network's own sence of relative position, which arises in a kinesthetic linkage between self and world.
Including the environment in the learning process is an increasingly important paradigm. Instead of being guided by an algorithmic set of rules, a learning entity is simply immersed in an environment, such as a chessboard with an active opponent, in which it must discover appropriate ways to behave. Learning and practice are combined; there is no longer any artificial distinction between training and performing. In the still-young discipline of designing autonomous machinery, Kuperstein's INFANT project appears overly ambitious. Better techniques are needed for balancing the conflicting requirements and plasticity and robustness, for assimilating new material while retaining the old.
Поделиться52012-05-03 20:13:12
Furthermore, connectionist implementations that learn only through weight adjustment of already-established connections adjust slowly to a charging environment, and seem inappropriate for modeling concepts that undergo radical change as exceptions are perceived.
Many of us see the genetic algorithm as a much more promising mechanism for implementing machine intuition. For one thing, the chess player examining game continuations, in search of a plan acts out the genetic algorithm. Over several generations - analysis cycles - the player examines a population of actions that seem appropriate in that board situation. Less fit, flawed continuations, if not discarded outright, are categorized as "remotely relevant" and rarely reconsidered; those that seem suitable and reexamined. Dangers and opportunities noticed while exploring earlier continuations spring to mind when relevant, so what initially distinct thoughts meet and fuse to produce new understanding.
This is exactly the behavior that an intuitive machine should exhibit. We would like to develop a contrivance that forms its own chess concepts through active exploration, and, like the human chess player, continually discovers new structure in a configuration of pieces.