Page 197 - Deep Learning
P. 197

180                         Adaptation

                                                          17
            behavior controlled by a goal as the scope of that goal.  A goal that controls
            behavior over a long time period is likely to be accomplished by calling upon
            goals of smaller scope rather than by calling on elementary actions directly.
            Such subordinate goals are called subgoals, and the dominant goal is their
            superordinate goal. A superordinate goal can be either a top goal or a subgoal
            to some other goal of yet greater scope. tracing the relations upward, make
            tea might turn out to be part of throw a dinner party. There is no upper
            limit on the scope of goals; consider achieve world peace. The goals that are
            involved  in  performing a given  task  jointly  form a goal  hierarchy or  goal
            tree. 18
               typically, all the subgoals to a given goal have to be accomplished for that
            goal to be accomplished. to make tea, you have to boil the water and put tea in
            the strainer and pour the water through the strainer. These subgoals are linked
            by  conjunction  (“and”).  From  the  point  of  view  of  the  superordinate  goal,  it
            might not matter exactly how any one subgoal is accomplished. A subgoal can
            be associated with different ways of achieving it; you can boil water over an open
            fire or on an electric store or in a microwave oven. These methods are alterna-
            tives. There is no need to boil the water three times; once will do. Similarly, go to
            dinner might entail tell X that we”ll meet him at the restaurant, but success in this
            part of the dinner project does not depend on whether X is told face-to-face or
            via fax, phone, pigeon post or texting; all that matters is that X is told, somehow.
            Alternative methods for achieving a goal are linked by disjunction (“or”). A goal
            tree consists of alternating conjunctive and disjunctive layers; multiple subgoals,
            each reachable by alternative paths, each of which in turn poses multiple sub-
            subgoals, and so on. in everyday parlance, we refer to such a structure as a plan;
            see Figure 6.2. The claim is that battle plans, business plans, lesson plans and
            travel plans all share these structural features. The computer science term AND/
            OR tree is awkward but descriptive of these complex structures.


                                    task Environments

            When a person decides to pursue a particular goal he is in the initial state of
            his performance. A situation in which the criteria or constraints that define the
            goal are satisfied is a goal state. Although some tasks can be accomplished with
            a single action – to turn on the light, flip the switch – it is more common that
            a task requires a sequence of actions. To open the window might require us to
            draw the curtain, lift the latch, push the window outward, and secure the latch
            again. Each action changes the material situation in some respect. A sequence
            of actions A , A , A , …, A  creates a sequence of situations, S , S , S  …, S . Each
                                                               2
                                                             1
                                                                       n
                                                                  3
                                 n
                           3
                        2
                     1
   192   193   194   195   196   197   198   199   200   201   202