5.2 Learned Reactions

All animals are born with ‘instincts’ like ‘get away from a quickly approaching object.’ Such built-in reactions tend to serve well so long as those animals stay in environments like those in which their instincts evolved. But when those worlds change, those creatures may need to be able to learn new ways to react. For example, when Joan perceives that oncoming car, she partly reacts instinctively, but she also depends on what she has learned about that particular kind of danger or threat. But how and what did she actually learn? We’ll come back to this toward the end of this book, because human learning is extremely complex, and here we’ll merely mention some ideas about how learning might work in some animals. During the 20th century, many well-known psychologists adopted this portrayal of how animals learn new If–>Do rules:

When an animal faces a new situation, it tries a random sequence of actions. Then, if one of these is followed by some ‘reward,’ then that reaction gets ‘reinforced.’ This makes that reaction more likely to happen when that animal faces the same situation.

This theory of ‘learning by reinforcement’ can be made to explain a good deal of what many kinds of animals do. Indeed, that theory was largely based on experiments with mice and rats, pigeons, dogs and cats, and snails. However, it does not help much to explain how people learn to solve difficult problems that require more complex series of actions. Indeed, deciding what to learn from these may be harder than actually solving those problems, and words like random, reward, and reinforce do not help us answer this two crucial questions:

How were the successful reactions produced? To solve a hard problem, one usually needs an intricate sequence of actions in which each step depends on what others have done. A lucky guess might produce one such step, but random choices would take far too long to find an effective sequence of them. We’ll discuss this below in Searching And Planning.

Which aspects of recent events to remember? For an If to work well, it must include only the relevant features, because one can be misled by irrelevant ones. (If you learned a new way to tie a knot, your Ifs should not mention the day of the week.) For as we’ll see in §8 Resourcefulness, if your description is too specific, then it will rarely match new situations—but if your description is too abstract, then it will match too many of them—and in either case, you won’t learn enough.

For example, suppose that you want a robot to recognize the visual image of any human hand. This is hard because we never see the same image twice—even of the very same hand—because each finger may change its position and shape, we’ll see it from different points of view, and each part will catch different amounts of light. This means that we’ll need trillions of If–>Do rules, unless we can find some special tricks that single out just the most relevant features—or if, as we’ll see in §6-2, we can formulate high-level descriptions like “a palm-shaped object with fingers attached.”

Certainly, many things that we do are based on reacting to external events by using simple If–>Do rules. However, along with those low-level reactions, we are also always making new plans and thinking about what we’ve done in the past—and those internal mental activities are what give us our unique abilities.

For example, when Joan reacted to that moving car, her reaction was partly instinctive and partly learned. However, she could not have ‘learned from experience’ that cars are especially dangerous—because if she had learned this by trial and error, she probably would not be alive; learning by ‘reinforcing’ success is a really bad way to learn to survive. Instead, she either ‘figured this out’ for herself or was told about it by someone else, and both of these must have involved higher levels of mental activities. So now let’s turn to what we call ‘thinking’—that is, the techniques that we use when we react, not just to events in the outer world, but also to other events in our brains.