当前位置：文档库 › Learning hierarchical task networks by observation

Learning hierarchical task networks by observation

Learning Hierarchical Task Networks by Observation

Negin Nejati negin@https://www.wendangku.net/doc/6a84256.html, Pat Langley langley@https://www.wendangku.net/doc/6a84256.html, Tolga Konik konik@https://www.wendangku.net/doc/6a84256.html, Computational Learning Laboratory,Center for the Study of Language and Information,Stanford University, Stanford,CA94305USA

Abstract

Knowledge-based planning methods o?er

bene?ts over classical techniques,but they

are time consuming and costly to construct.

There has been research on learning plan

knowledge from search,but this can take sub-

stantial computer time and may even fail

to?nd solutions on complex tasks.Here

we describe another approach that observes

sequences of operators taken from expert

solutions to problems and learns hierarchi-

cal task networks from them.The method

has similarities to previous algorithms for

explanation-based learning,but di?ers in its

ability to acquire hierarchical structures and

in the generality of learned conditions.These

increase the method’s capability to trans-

fer learned knowledge to other problems and

supports the acquisition of recursive proce-

dures.After presenting the learning algo-

rithm,we report experiments that compare

its abilities to other techniques on two plan-

ning domains.In closing,we review related

work and directions for future research.

1.Introduction

In recent years,hierarchical task networks have emerged as a powerful framework for representing and organizing knowledge about action(Ilghami et al., 2002;Wilkins&desJardins,2001).With access to such content,an agent can generate or execute plans far more e?ectively than it can from domain operators alone,because the knowledge speci?es how to decom-pose complex tasks into simpler ones.Such approaches have been applied successfully to a variety of challeng-Appearing in Proceedings of the23rd International Con-ference on Machine Learning,Pittsburgh,PA,2006.Copy-right2006by the authors.ing domains,and they scale to complexity much better than classical planning methods.

Despite their clear advantages,hierarchical task net-works can be di?cult and time consuming to construct manually.This suggests the use of machine learning to acquire the knowledge base from experience,but,de-spite a substantial literature on learning for planning (Zimmerman&Kambhampati,2003),there have been remarkably few e?orts toward acquiring hierarchical plan structures.Moreover,most work in this area has focused on learning from the results of search,which can produce poor solutions and may even fail entirely on some tasks.

An alternative approach is to learn from traces of an expert’s behavior.Research on behavioral cloning (e.g.,Sammut,1996)incorporates this idea,but typ-ically learns?at rules for reactive control.Two other paradigms–learning apprentices(e.g.,Mitchell et al.,1985)and programming by demonstration(e.g., Cypher,1993)–emphasize di?erent induction meth-ods and problems but also focus on nonhierarchical structures.In this paper,we adapt the general idea of learning from expert traces to acquire a class of hi-erarchical task networks–teleoreactive logic programs –that are distinctive because they index methods by the goals they achieve.We can state the learning task more precisely in terms of its inputs and outputs:?Given:a set of operators that produce predictable e?ects under known conditions;

?Given:a set of training problems,each of which speci?es an initial state and a goal;

?Given:for each training problem,a sequence of operator instances that achieves the goal from the initial state;

?Find:a teleoreactive logic program that repro-duces the solutions to the training problems and generalizes well to new ones.

We will see that,unlike research on behavioral cloning, our approach chains backward from the goal achieved by each sample trace to analyze the reasons for each action.In this sense,it bears similarity to early work on explanation-based learning,which also relied on goal regression.However,these methods jettisoned the explanation structure after learning,whereas our approach retains it to determine the hierarchical orga-nization of learned knowledge.

In the next section,we describe the knowledge struc-tures given to and acquired by our method,along with the performance mechanisms that operate over that knowledge.After this,we describe the learning algo-rithm in detail and illustrate its operation on a simple example.Next we report experiments with the tech-nique on two planning domains,including comparisons with more traditional learning methods.Finally,we examine at more length how our approach relates to earlier work on learning plan knowledge and discuss some directions for additional research.

2.Knowledge Representation

Before we can explain our performance and learning methods,we must?rst describe the representation we use to encode the knowledge they use and acquire. This formalism supports a specialized class of hier-archical task networks that we call teleoreactive logic programs,which Langley and Choi(2006)report in more detail.Such programs consist of two distinct but interconnected knowledge bases that describe di?erent aspects of a domain.

The?rst knowledge base includes a set of concept def-initions that recognize classes of situations in the do-main and describe them at di?erent levels of abstrac-tion.Concepts are encoded in a hierarchical language and shape the system’s beliefs about the domain.Each concept includes a head,which is stated as a predicate with zero or more arguments,and a body,which in-cludes one or more arithmetic tests or relations that must hold.The lowest-level concepts only refer to the perceptual inputs from the domain,while the higher level ones can refer to other concepts as well.Table1 shows examples of two primitive concept de?nitions, on and hand-empty,and two nonprimitive concepts, clear and unstackable,for the Blocks World.

The second knowledge base contains the de?nitions of primitive and high-level skills that can be executed to change the environment.Primitive skills are de?ned in terms of the actions the agent can perform with each clause having a head that speci?es its name and argu-ments,a set of typed variables,a single start condition, a set of e?ects,and a set of executable actions.The Table1.Examples of concepts from the Blocks World. (on(?blk1?blk2)

:percepts((block?blk1x?x1y?y1)

(block?blk2x?x2y?y2h?h))

:tests((equal?x1?x2)

(>=?y1?y2)

(<=?y1(+?y2?h))))

(clear(?block)

:percepts((block?block))

:negatives((on?other?block)))

(unstackable(?block?from)

:percepts((block?block)(block?from))

:positives((on?block?from)

(clear?block)(hand-empty))) (hand-empty()

:percepts((hand?hand status?status))

:tests((eq?status’empty)))

e?ects of each primitive skill are known to the system and are speci?ed in terms of known concepts.These correspond to primitive tasks or operators in hierarchi-cal task networks.

Higher level skills,which are the result of the learn-ing,specify ways to decompose complex activities into simple ones.They correspond to methods in hierarchi-cal task networks but di?er in that the head of each skill clause is the goal concept it achieves if executed to completion.This introduces a connection between concepts and skills that is essential for our learning method,as we describe in Section4.The body of a high-level skill includes the necessary perceptual in-puts,the start conditions,and the relevant subskills. Table2and Figure1show examples of primitive and high-level skills from the Blocks World.

Both concept and skill knowledge bases are stored in long-term memories,while the speci?c instances of them are stored in short-term memories that change as the environment evolves.

3.Inference and Execution Mechanisms The performance mechanisms of a teleoreactive logic program re?ect the fact that it operates in a physi-cal setting that changes over time.As Langley and Choi(2006)report,the basic architecture functions in discrete cycles.On each cycle it deposits,in a per-ceptual bu?er,low-level sensory information about ob-jects within the agent’s?eld of view.The system then calls an inference module that operates in a bottom-up,data-driven manner to infer relational beliefs from these perceptions and concept de?nitions.The in-ference procedure?rst produces instances of primi-

Table2.Two primitive skills for the Blocks World domain. (unstack(?block?from)

:percepts((block?block ypos?ypos)

(block?from))

:start((unstackable?block?from))

:effects((clear?from)

(holding?block))

:actions((*grasp?block)

(*v-move?block(+?ypos10)))) (stack(?block?to)

:percepts((block?block)

(block?to x?x y?y h?h))

:start((stackable?block?to))

:effects((on?block?to)

(hand-empty))

:actions((*h-move?block?x)

(*v-move?block(+?y?h))

(*ungrasp?block)))

tive concepts,which depend only on perceived objects, then asserts beliefs based on higher-level concepts. After this,the architecture invokes an execution mod-ule that,in contrast with inference,operates in a top-down manner to achieve high-level goals.On each cy-cle,using the beliefs produced during inference,the system?nds all applicable paths through the skill hi-erarchy that start from the agent’s goal.Since goals are stated as concept instances and the heads of non-primitive skills refer to the concept they achieve,the agent can easily pick executable paths that are rele-vant to its current goal.The applicability of a skill path depends on the applicability of every skill clause on that path.An individual skill instance is appli-cable when in the current state its preconditions are satis?ed but its head is not satis?ed.Note that a skill path is not about a course of action over time;it de-scribes the hierarchical context from the highest-level skill down to a primitive skill.When the agent carries out the actions associated with the selected skill path, the environment changes and the architecture repeats the same procedure until it achieves the goal.

To assure reactivity,on the cycle the system also con-siders the previously executed subskills.This enables it to invoke those subskills again if their objectives are no longer satis?ed due to unexpected events.On the other hand,to prevent the agent from switching among multiple intentions instead of persistently working on one and moving to the next,the architecture prefers skill paths that are more similar to the path that it chose on the previous cycle.

Our approach to skill utilization bears many similari-ties to those for hierarchical task networks(Wilkins&Table3.A training example for learning by observation from the Blocks World.

Goal:(clear A)

Primitive skill sequence:

((unstack C B)(putdown C T1)(unstack B A)) Initial state:

((unstackable C B)(hand-empty)(clear C)

(ontable A T1)(on C B)(on B A)) desJardins,2001),which have often emphasized plan execution over plan generation.However,most work in this tradition has relied on sophisticated notations for describing complex activities that make them dif-?cult to learn.Our formalism for teleoreactive logic programs is closer to the one assumed by Ilghami et al.’s(2002)SHOP2,which focuses on plan gener-ation rather than execution.As we will see shortly, this representation supports the incremental learning of hierarchical skills from traces of expert behavior.

4.Learning Mechanism

Now that we have described the processes that inter-pret a teleoreactive logic program to achieve an agent’s goals,we can turn to methods for learning these struc-tures by observation.As mentioned earlier,our ap-proach builds on knowledge about a domain and ex-pands upon it to incorporate skills that can achieve complicated tasks.This background knowledge in-cludes the de?nitions for concepts and primitive skills. Our system learns from problem-solution pairs pro-vided by the expert.Each problem is de?ned by a goal instance and a speci?c initial state of the envi-ronment.The solution is a sequence of primitive skill instances provided by the expert that achieves the goal starting from the initial state.Table3shows a train-ing example for a simple problem in the Blocks World domain.The initial state for this problem has a three-block tower with C on B and B on A,and the goal is to get block A clear.

Given these inputs,the algorithm proceeds to parse and explain the solution steps using a strategy similar to that in explanation-based learning(Ellman,1989). It starts by chaining backward from the goal by focus-ing on the?nal primitive skill provided by the expert, such as(unstack B A)in our example from Table3. At each step,it selects between two alternative ac-counts of the expert trace.

If the primitive skill contains the goal as one of its ef-fects,the algorithm explains the goal using skill chain-ing.If this skill is applicable in the initial state of

the problem,one can directly execute it to achieve the goal.However,if its precondition is not satis?ed,extra work is needed to make it applicable.In this case,the algorithm tags the single precondition of this skill as its new goal and tries to justify the previous solution steps with respect to it.Figure1illustrates the parsing procedure on the Blocks World example from Table3. In this example,the last primitive skill,(unstack B A),achieves the goal(clear A)directly,but its pre-condition,(unstackable B A),is not satis?ed in the initial state.This unsatis?ed precondition is consid-ered as the goal the expert was pursuing before the?-nal move.Therefore,the algorithm proceeds to justify the primitive skill sequence(unstack C B)(putdown C T1)for achieving(unstackable B A).

When the goal is not one of the e?ects of the prim-itive skill in the sequence,the algorithm explains it using concept chaining.In this case,the skill either achieves one of the goal’s subconcepts or it is an un-necessary step that presumably achieves some other purpose.The system determines the possible subgoals from the de?nition of the goal concept,and then looks back through the state sequence to?nd the order in which the expert achieved them.All primitive skills executed before achieving a subgoal are treated as po-tential contributors.This can lead to considering a skill for more than one subgoal,which is reasonable since skills can have multiple e?ects that contribute to di?erent aims.In addition,this lets the system handle interleaved subtraces when the expert alternates be-tween pursuing di?erent subgoals.When a subtrace includes a skill that is irrelevant to achieving a goal, the system simply ignores it.Each subgoal with its as-sociated sequence of primitive skills poses a new prob-lem is handled by calling the main routine recursively. Continuing with our Blocks World example,the sequence(unstack C B)(putdown C)in Table3 makes B and A unstackable,but none of these prim-itive skills include(unstackable B A)as their ef-fects.This results in concept chaining and break-ing down(unstackable B A)into its subgoals–(hand-empty),(on B A),and(clear B)–as de-?ned in Table1.By tracking the order of subgoal achievements through the state sequence,(unstack C B)is assigned to(clear B),while both(unstack C B)and(putdown C)are considered as contributors to achieving(hand-empty).The method will detect the literal(unstack C B)as an irrelevant step for this goal later in its processing.

After the algorithm has?nished parsing the expert’s solution trace,it uses the explanation structure to learn new skill clauses for future use.For skill chain-ing,if the precondition P of the current primitive

skill Figure1.Illustration of the learning algorithm on a simple problem from the Blocks World and the acquired skills.

S is satis?ed in the initial state and the expert did nothing to achieve that,the system learns a new skill clause with the achieved goal G as its head,P as its start condition,and a single subskill S.1If the ex-pert has done work to achieve P,the system learns a skill clause that achieves G by composing the skill that achieves P(which has the same head as P and has already been learned in the shallower recursion calls) with S.The start conditions of the learned skill are the same as the start conditions of the skill P.In Fig-ure1,(clear?B)and(clear?A)are examples of skill clauses learned through skill chaining.

For concept chaining,if the goal G has n subgoals, g1,...,g n,and m of them are satis?ed in the initial state,the lower level skills that are learned for achiev-ing g m+1,...,g n comprise the subskills?eld of the new skill,but the g1,...,g m concepts determine the start condition of this new skill.The nonprimitive skill (unstackable?B?A)in Figure1is an example of a skill clause learned by this procedure.As we discuss shortly,the acquired precondition does not guarantee successful execution,but it still often serves as a use-ful heuristic.The learning mechanism can produce disjunctive skills,since di?erent methods of achieving a goal have the same head.This convention leads to ?exible skills,since each clause can be expanded with all combinations of its subskills’di?erent clauses.

As demonstrated elsewhere(Langley&Choi,2006), one can use traces generated by a problem solver to learn analogous skills.The method introduced here replaces the problem-solving traces with ones from an expert performing the task.Although we use a simi-lar subroutine for creating skill clauses,the resulting hierarchy can di?er because the expert may provide di?erent solution traces than the problem solver.We 1For the distinction between this learned clause and the primitive skill it incorporates,see Langley and Choi(2006).

expect that learning by observation will be more prac-tical in complex domains because of the heavy search they impose on problem solvers.

5.Experimental Evaluation

To evaluate the behavior of our approach to learning teleoreactive logic programs by observation,we de-signed and carried out experiments in two planning domains.Below we describe our experimental design, after which we report the results of these studies.

5.1.Experimental Design

Since our method operates in an incremental man-ner,we chose to employ an online training regimen in which we presented randomly sampled problems one at a time.If the system fails to achieve the goal using its current knowledge,it retrieves an expert trace stored with the problem.After learning a set of skill clauses from this trace,the system turns to the next problem. This lets us measure performance as a function of the number of problems encountered.We selected the cu-mulative number of tasks solved successfully as our performance metric,primarily because di?erent prob-lems involved distinct levels of di?culty,and thus were not directly comparable.We also averaged the results across di?erent sequences of problems to guard against order e?ects.Other metrics such as CPU time are not relevant here,as the performance system does not in-clude a problem solver and simply fails to solve a task if it does not have all the necessary skills.

Because our approach to learning by observation con-structs an explanation of the expert trace,it seems appropriate that we compare its behavior to tradi-tional methods for explanation-based learning.We considered two variations on this idea,each tied to one of our theoretical claims.First,we maintain that our method should acquire more?exible knowledge than explanation-based techniques because it retains the explanation structure in its skill hierarchy.Thus, we implemented a standard algorithm for learning a ?at macro-operator(Mooney,1990)from each expert trace that collects a set of su?cient conditions,with constants replaced by variables,to achieve that prob-lem’s top-level goal.This macro-operator can solve the training problem on which it is based,but the method ?nds very speci?c conditions,and preliminary experi-ments suggested it generalized so poorly to new tasks that we did not pursue it further.

A second claim revolves around our method for select-ing conditions on learned skills,which does not rely on goal regression and only uses information about component skills and concepts.For comparison pur-poses,we developed an alternative technique that still learns hierarchical skills by analyzing expert traces, but that invokes the same condition-?nding scheme for each clause as for macro-operator formation.This produces conditions that are su?cient to achieve the goal in each clause’s head,making them more speci?c than those our method generates,but also less likely to initiate chains of action that the agent cannot com-plete.Initial studies suggested that this alternative generalized reasonably well,so we included it in our formal experiments.

We settled on two domains for our studies:the Blocks World and Depots,which we describe more fully later. The former is not especially di?cult,but it is well understood,and we had a handcrafted task network that we could compare with learned skills.Depots is a more challenging domain that involves a much larger search space,which we predicted would help distinguish our approach from the explanation-based alternative.However,because we had no handcrafted program against which to compare learned skills,we could only evaluate the methods on this domain in terms of their behavior.

5.2.Experiments with the Blocks World

The Blocks World consists of a table,a gripper,and several cubical blocks that are distributed in one or multiple columns on the table.The system distin-guishes di?erent situations in the environment using nine de?ned concepts,including the ones in Table1. It can also change its environment using four primi-tive skills:stack,unstack(see Table2),pickup,and putdown.The goals refer to con?gurations of blocks that are de?ned via existing concepts.This knowledge is su?cient,in principle,to solve all the problems in the domain by employing extensive search.However, when there are more than a few blocks in the environ-ment,search is slow and impractical.

Our experiment used240random problems with three distinct goals:clear,holding,and on.The environ-ment included up to seven blocks with random initial con?gurations and the di?culty level of the problems ranged from one to16-step solutions.Inspection re-vealed the system learned skills that are equivalent to the handcrafted ones.Figure1shows the skills acquired from the sample input in Table3.These high-level skills equip the system to achieve goals like clearing a block on many new problems.One reason is their potential for being applied recursively.For example,the learned skills in the?gure can clear a block regardless of the number of blocks on top of

Table4.Some examples of skills acquired in the Blocks World using explanation-based learning.

(clear(?A)

:percepts((block?A)(block?B)

(block?C))

:start((hand-empty)(on?B?A)

(on?C?B)(clear?C)

:ordered((unstackable?B?A)

(unstack?B?A)))

(unstackable(?B?A)

:percepts((block?A)(block?B)

(block?C))

:start((hand-empty)(on?B?A)

(on?C?B)(clear?C))

:ordered((clear?B)(hand-empty)))

it.Table4shows the analogous skills learned by the explanation-based method,which have very speci?c start conditions.This results in learning distinct ver-sions of(clear?A)for problems with di?erent num-ber of blocks stacked on block A.

The left hand graph in Figure2shows the superior performance of our system compared to explanation-based learning.The plots show the cumulative num-ber of successful examples averaged over20di?erent selections and orderings of100input problems in the Blocks World domain for each algorithm.On average, our system learns14new skills from a few problems and solves100percent of the test problems,while the explanation-based method learns119skills and solves fewer than50percent of the novel problems.The dot-ted line shows the behavior of an expert who can solve all problems from the outset.Our method’s behaviour follows this line very closely,with a small lag caused by a few failures at the early stages of learning.The small error bars show that the behaviors are similar across di?erent problem orderings.

5.3.Experiments with Depots

Depots is a more complicated domain that was intro-duced in the Third International Planning Competi-tion(Bacchus,2000).With crates that can be loaded into trucks and driven to di?erent locations where they are unloaded and stacked onto pallets,it combines at-tributes of Blocks World with logistics planning.Since a typical problem involves many objects and each state has many possible actions,search can be very expen-sive in this domain and manually coding the knowledge base is challenging.We gave the system23concepts and eight primitive skills,and we used two di?erent goals that refer to con?gurations of one or two crates.For this study,we generated96random problems for use in the experiment.These problems include up to ?ve crates,with di?culty levels ranging from one to 13-step solution paths depending on the initial con-?guration.The right-hand graph in Figure2shows the cumulative number of successes averaged over20 di?erent problem orderings for our algorithm.

The number of skills our system learns through this experiment is42on average,which is not su?cient to solve all the problems,but which does solve70per-cent of them.For this domain,although the hierarchi-cal explanation-based algorithm generated some cor-rect skills,it was very slow and could solve very few of the problems.As anticipated,because the learned skills are overspeci?c,it creates many skills(e.g.,some 100skills resulted from only four problems).However, we may be able to improve these results by generat-ing multiple traces from a single problem.For ex-ample,if a goal G is associated with a skill sequence S1,S2,S3,...,S n,the current technique learns one skill that achieves G using these subskills,but the subse-quences(S2,S3,...,S n),(S3,S4,...,S n),...,(S n)are also valid sequences for achieving G and can be used for training purposes.On the other hand,this would produce a more complex collection of skills that could slow response time further.

6.Related Research

Developing agents that can learn complex skills from experience has been a recurring goal in arti?cial in-telligence.A common approach to this problem has focused on learning from delayed external rewards. Some methods(e.g.,Moriarty et al.,1999)search through the space of the policies directly,whereas others(e.g.,Kaelbling et al.,1996)estimate value functions for state-action pairs.In contrast,our ap-proach di?ers by learning from traces of expert behav-ior rather than from exploration and by constructing hierarchical structures rather than?at policies. There has been some work on learning control poli-cies from expert traces.One of the main paradigms, known as behavioral cloning,transforms the observed traces into supervised training cases and induces re-active controllers that reproduce the behavior in sim-ilar situations.This approach typically casts learned knowledge as decision trees that determine which ac-tions to take based on sensor input(e.g.,Urbancic& Bratko,1994).More recently,some e?orts(e.g.,Isaac &Sammut,2003)have used goal information,stated as desired values for state variables,to improve the robustness of the learned controllers.In contrast,our approach infers intermediate structural goals by ex-

02040

6080100

Number of problems encountered

20406080100

C u m u l a t i v e p r o b l e m s s o l v e d Expert behavior

Explanation-based learning Hier. task network learning 600

2040

80100Number of problems encountered

2040

6080100

C u m u l a t i v e p r o b l e m s s o l v e d

Expert behavior

Hier. task network learning Figure 2.Cumulative number of solved problems in the Blocks World and Depots domains,with 95percent con?dence intervals.The diagonal lines depict the behavior of an expert who solves all the problems.

plaining the expert behavior in terms of the top-level goal and background knowledge.Moreover,our frame-work di?ers from most work on behavioral cloning by incorporating a hierarchical and relational representa-tion for states and activities.

K¨o nik and Laird (2004)report an approach to behav-ioral cloning that uses a relational representation and learns hierarchical controllers,but it requires the ex-pert to annotate the trace with information about the start and end of activities at each level.A few other systems also learn hierarchical structures from expert behavior.Ilghami et al.(2005)describe a method for constructing hierarchical task networks,but they assume the hierarchical structure is given and use a version-space algorithm to determine conditions on methods.Tecuci’s (1998)Disciple acquires hierarchi-cal rules,but it requires user information about how to decompose problems and the reasons for decisions.Other research on learning skills by observing others’behavior has,like our own,utilized domain knowledge to interpret traces.Some work on explanation-based learning (e.g.,Segre,1987;Mooney 1990)took this approach,as did the paradigms of learning appren-tices (e.g.,Mitchell et al.,1985)and programming by demonstration (e.g.,Cypher,1993;Lau et al.,2003).Both often used analytic methods to generate candi-date procedures,but neither focused on the acquisi-tion of hierarchical skills,and programming by demon-stration typically requires user feedback about candi-date hypotheses.As noted earlier,our approach dif-fers from explanation-based learning in that it retains the explanation structure and does not use deductive analysis to determine start conditions on new skills.

7.Concluding Remarks

In this paper,we have presented a new approach to learning hierarchical task networks from observation.We explained our reliance on expert traces to avoid the need for extensive search during problem solving.We also reviewed the notion of teleoreactive logic pro-grams,a special class of task networks that are indexed by the goals they achieve,along with mechanisms for executing them to reach these objectives.After this,we described the details of our method for learning skill clauses from solution traces,which alternates between chaining o?skills and concept de?nitions to generate an explanation of the expert’s behavior.The system then creates one skill clause for each node in the expla-nation structure,using a simple technique for ?nding conditions on clauses that uses only local information.Our approach incorporates ideas from a number of dis-tinct traditions,including behavioral cloning,learn-ing apprentices,and programming by demonstration,but it goes beyond these movements to construct hi-erarchical structures that generalize well to new prob-lems.Moreover,experiments in two domains demon-strated that our method learns much more rapidly than explanation-based techniques,whether they con-struct ?at macro-operators or take advantage of the learned skill hierarchy.This suggests that our ap-proach to constructing hierarchical task networks of-fers advantages over more traditional techniques for learning plan knowledge.

Despite these encouraging results,our work on learn-ing by observation remains in its early stages,and our current implementation makes some simplifying as-

sumptions that we should remedy in future work.One assumption is that the expert trace includes all rele-vant steps,whereas a more robust agent might infer oc-casional steps that it does not observe.Another is that learner knows the agent’s goals,whereas an improved system would infer them from the trace and back-ground knowledge.Our current method also assumes that the environment changes only when the agent takes some action,whereas future versions should han-dle settings in which some events are due to indepen-dent physical causes or other agents’behaviors.

In addition,our approach relies heavily on accurate knowledge about each skill’s e?ects,whereas a more ?exible learner would also include some ability to re-vise its action models based on observed results.Fi-nally,the current system assumes that the expert has carried out the best sequence to achieve the goal, whereas a more sophisticated approach would analyze the trace for loops or other ine?ciencies and remove them before learning.We hope to address each of these issues in future learning systems that combine domain knowledge with traces of expert behavior to acquire complex hierarchical task networks. Acknowledgements

This paper reports research sponsored by DARPA un-der agreement FA8750-05-2-0283.The https://www.wendangku.net/doc/6a84256.html,ern-ment may reproduce and distribute reprints for Gov-ernmental purposes notwithstanding any copyrights. The authors’views and conclusions should not be in-terpreted as representing o?cial policies or endorse-ments,expressed on implied,of DARPA or the Gov-ernment.We thank Dongkyu Choi for discussions that contributed to the ideas presented in this paper. References

Bacchus,F.(2001).AIPS’00planning competition. AI Magazine,22,47–56.

Cypher,A.(Ed.).(1993).Watch what I do:Program-ming by demonstration.Cambridge,MA:MIT Press. Ellman,T.(1989).Explanation-based learning:A sur-vey of programs and perspectives.ACM Computing Surveys,21,163–221.

Ilghami,O.,Nau,D.S.,Mu?n oz-Avila,H.,&Aha,D. W.(2002).CaMeL:Learning method preconditions for HTN planning.Proceedings of the Sixth Interna-tional Conference on AI Planning and Scheduling(pp. 131–14).Toulouse,France.Kaelbling,L.P.,Littman,L.M.,&Moore,A.W. (1996).Reinforcement learning:A survey.Journal of Arti?cial Intelligence Research,4,237–285.

K¨o nik,T.,&Laird,J.(2004).Learning goal hier-archies from structured observations and expert an-notations.Proceedings of the Fourteenth International Conference on Inductive Logic Programming(pp.198–215).Porto,Portugal:Springer.

Langley,P.,&Choi,D.(2006)Learning recursive con-trol programs from problem solving.Journal of Ma-chine Learning Research,7,493–518.

Lau,T. A.,Domingos,P.,&Weld, D.S.(2003). Learning programs from traces using version space al-gebra.Proceedings of the Second International Con-ference on Knowledge Capture(pp.36–43).Sanibel Island,FL:ACM Press.

Mitchell,T.M.,Mahadevan,S.,&Steinberg,L.I. (1985).LEAP:A learning apprentice for VLSI design. Proceedings of the Ninth International Joint Confer-ence on Arti?cial Intelligence(pp.573–580).Los An-geles,CA:Morgan Kaufmann.

Mooney,R.J.(1990).A general explanation-based learning mechanism and its application to narrative understanding.San Mateo,CA:Morgan Kaufmann. Moriarty,D.E.,Schultz,A.C.,&Grefenstette,J.J. (1999).Evolutionary algorithms for reinforcement learning.Journal of Arti?cial Intelligence Research, 11,241–276.

Sammut,C.(1996).Automatic construction of reac-tive control systems using symbolic machine learning. Knowledge Engineering Review,11,27–42.

Segre,A.(1987).A learning apprentice system for mechanical assembly.Proceedings of the Third IEEE Conference on AI for Applications(pp.112–117). Tecuci,G.(1998).Building intelligent agents:An apprenticeship multistrategy learning theory,method-ology,tool and case studies.London:Academic Press. Urbancic,T.,&Bratko,I.(1994).Reconstructing human skill with machine learning.Proceedings of the Eleventh European Conference on Arti?cial Intel-ligence(pp.498–502).Amsterdam:John Wiley. Wilkins,D.E.,&desJardins,M.(2001).A call for knowledge-based planning.AI Magazine,22,99–115. Zimmerman,T.,&Kambhampati,S.(2003).Learn-ing-assisted automated planning:Looking back,tak-ing stock,going forward.AI Magazine,24,73–96.

新版人教版高中语文课本的目录。

必修一阅读鉴赏第一单元1.*沁园春?长沙……………………………………毛泽东3 2.诗两首雨巷…………………………………………戴望舒6 再别康桥………………………………………徐志摩8 3.大堰河--我的保姆………………………………艾青10 第二单元4.烛之武退秦师………………………………….《左传》16 5.荆轲刺秦王………………………………….《战国策》18 6.*鸿门宴……………………………………..司马迁22 第三单元7.记念刘和珍君……………………………………鲁迅27 8.小狗包弟……………………………………….巴金32 9.*记梁任公先生的一次演讲…………………………梁实秋36 第四单元10.短新闻两篇别了，“不列颠尼亚”…………………………周婷杨兴39 奥斯维辛没有什么新闻…………………………罗森塔尔41 11.包身工………………………………………..夏衍44 12.*飞向太空的航程……………………….贾永曹智白瑞雪52 必修二阅读鉴赏第一单元1.荷塘月色…………………………………..朱自清2.故都的秋…………………………………..郁达夫3.*囚绿记…………………………………..陆蠡第二单元4.《诗经》两首氓采薇5.离骚………………………………………屈原6.*《孔雀东南飞》(并序) 7.*诗三首涉江采芙蓉《古诗十九首》短歌行……………………………………曹操归园田居(其一)…………………………..陶渊明第三单元8.兰亭集序……………………………………王羲之9.赤壁赋……………………………………..苏轼10.*游褒禅山记………………………………王安石第四单元11.就任北京大学校长之演说……………………..蔡元培12.我有一个梦想………………………………马丁?路德?金1 3.*在马克思墓前的讲话…………………………恩格斯第三册阅读鉴赏第一单元1.林黛玉进贾府………………………………….曹雪芹2.祝福………………………………………..鲁迅3. *老人与海…………………………………….海明威第二单元4.蜀道难……………………………………….李白5.杜甫诗三首秋兴八首(其一) 咏怀古迹(其三) 登高6.琵琶行(并序)………………………………..白居易7.*李商隐诗两首锦瑟马嵬(其二) 第三单元8.寡人之于国也…………………………………《孟子》9.劝学……………………………………….《荀子》10.*过秦论…………………………………….贾谊11.*师说………………………………………韩愈第四单元12.动物游戏之谜………………………………..周立明13.宇宙的边疆………………………………….卡尔?萨根14.*凤蝶外传……………………………………董纯才15.*一名物理学家的教育历程……………………….加来道雄第四册阅读鉴赏第一单元1.窦娥冤………………………………………..关汉卿2.雷雨………………………………………….曹禹3.*哈姆莱特……………………………………莎士比亚第二单元4.柳永词两首望海潮(东南形胜) 雨霖铃(寒蝉凄切) 5.苏轼词两首念奴娇?赤壁怀古定风波(莫听穿林打叶声) 6.辛弃疾词两首水龙吟?登建康赏心亭永遇乐?京口北固亭怀古7.*李清照词两首醉花阴(薄雾浓云愁永昼) 声声慢(寻寻觅觅) 第三单元8.拿来主义……………………………………….鲁迅9.父母与孩子之间的爱……………………………..弗罗姆10.*短文三篇热爱生

托福口语task3-6

名师支招：新托福口语Task 3-6 完全攻略：征服新托福口语Task 3 “To succeed academically in English-speaking colleges and universities，students need to be able to combine all their English-language skills- in reading，listening，and speaking…”，所以在新托福口语task3中考生们首先遭遇了先读，再听，后说相结合的考试形式，虽然内容为较易理解的校园生活场景，但环环相扣的考试环节加上严格紧张的时间限制，不免给备考带来了种种压力。在本文中，笔者将根据Task 3 各个环节中经常出现的误区和问题进行探讨,并给出相应的解决方案，以助托福准考生们一臂之力。一．怎样“读”? 误区：阅读部分考察“记忆”能力相对于托福考试的阅读部分，Task3中只有75-100字的简短阅读材料简直是小巫见大巫。可正是因为这样，却偏偏激发了好多考生的“完美主义”心态，想要把所有内容统统记下，结果因小失大，笔记中过多的细节淹没了重点信息。其实Task3中阅读资料的真正作用是为听力提供背景，唯一的价值是告诉考生接下来的对话中两人谈论的政策或计划等是什么，所以不应浪费时间记过多细节。支招：笔记中只需要notice/message/proposal (key words) + reasons(keywords)。其中关于政策的关键词多半出现在题目以及文章的前两行，而通过一些明显的“信号”也可以轻松找到给出理由的关键词。例如：(Official Guide P213)announcement 关键词位于第一行“increase tuition”，而通过“but itis necessary to increase them now for several reasons”和“we havealso…”可以找到并列关系的两组理由关键词“more students，additionalprofessors”和“laboratoryfacilities”。所以笔记可记做：学费（tuition）↑，学生（students）+，老师（professors）+，实验设备（labfacilities）+(可根据习惯中英符号相结合)。二．怎样“听”? 误区1：对话中的两个人都听我们都知道在最终的口语表达中只需陈述其中一个人的态度及理由，所以两个说话者的信息都听无疑只会分散精力，并且另一个人所描述的信息也可能会被考生误引入到最后的口语表达中,违背题目要求给出错误信息。支招：找出主角（leading speaker），只听一人在听力对话一开始的时候不要着急记笔记，这样很容易无法分清主次。先把两个人的分别一句话听完，基本就可以分清主角和次角，因为，主角往往会持有很强烈的支持或反对态度，确定下主次以后，就可以专注只听一个人并记笔记了。误区2：听力对话考察“听写”能力有的考生曾经抱怨说自己的笔记都快赶上听力原文了，可是分数却不尽如人意。其实这又是“完美主义”心态在作怪了。考生总是不放心，对话中的每一个细节列举都不放过，结果在口语表达时内容冗长，却依然会遗漏重点信息，而这正是Task3丢分的主要原因之一。托福口语考察的是抓住重点的能力，并不是考生的书写速度。支招：笔记只需要记主角所给出的态度（赞成/反对）+reasons（key words）在听力对话中寻找关键信息，快速识别信号词或信号句就显得更为重要，这是听力的必备技巧之一。例如在(Official GuideP214-215)，我们可以看到“But I can see why”，“ And the other thing is ”，“I’mdoing OK, but the facilities here are so limited.”这些明显的信号句所引出的关键信息，所以笔记可记为：many students（班级+），hard get personalattention （关注?），not job（工作难），facilities?equipment ?（设备?）out of date（过时）三．怎样“说”? 误区：语速要快，尤其是在时间快结束时，这样才会多提供信息防止丢分。考生在时间来不及时慌慌张张讲得飞快，是因为想把笔记中的内容讲完，但在托福考官看来“how clearly and coherentlyyou convey information is as important as how much information youconvey”（表达的清晰与连贯和表达内容的多少是一样重要的），仓促作答一定顾不得清晰与连贯，所以这样的表现往往会失分。支招：备考时采用“计时录音”的方法进行练习。如果平时练习时就有时间的概念，那么考试时对口语表达的内容长短就比较有把握了，练习越多，就越胸有成竹。这个

连锁咖啡厅顾客满意度涉入程度对忠诚度影响之研究以高雄市星巴克为例

连锁咖啡厅顾客满意度涉入程度对忠诚度影响之研究以高雄市星巴克为例 Document number【SA80SAB-SAA9SYT-SAATC-SA6UT-SA18】

连锁咖啡厅顾客满意度对顾客忠诚度之影响－以高雄市星巴克为例 The Effects of Customer Satisfaction on Customer Loyalty—An Empirical study of Starbucks Coffee Stores in Kaohsiung City 吴明峻 Ming-Chun Wu 高雄应用科技大学观光管理系四观二甲学号：06 中文摘要本研究主要在探讨连锁咖啡厅顾客满意度对顾客忠诚度的影响。本研究以高雄市为研究地区，并选择8间星巴克连锁咖啡厅的顾客作为研究对象，问卷至2006年1月底回收完毕。本研究将顾客满意度分为五类，分别是咖啡、餐点、服务、咖啡厅内的设施与气氛和企业形象与整体价值感；将顾客忠诚度分为三类，分别是顾客再购意愿、向他人推荐意愿和价格容忍度，并使用李克特量表来进行测量。根据过往研究预期得知以下研究结果： 1.人口统计变项与消费型态变项有关。 2.人口统计变项与消费型态变项跟顾客满意度有关。 3.人口统计变项与消费型态变项跟顾客忠诚度有关。 4.顾客满意度对顾客忠诚度相互影响。关键词：连锁、咖啡厅、顾客满意度、顾客忠诚度 E-mail

一、绪论研究动机近年来，国内咖啡消费人口迅速增加，国外知名咖啡连锁品牌相继进入台湾，全都是因为看好国内咖啡消费市场。在国内较知名的连锁咖啡厅像是星巴克、西雅图极品等。本研究针对连锁咖啡厅之顾客满意度与顾客忠诚度间关系加以探讨。研究目的本研究所要探讨的是顾客满意度对顾客忠诚度的影响，以国内知名的连锁咖啡厅星巴克之顾客为研究对象。本研究有下列五项研究目的： 1.以星巴克为例，探讨连锁咖啡厅的顾客满意度对顾客忠诚度之影响。 2.以星巴克为例，探讨顾客满意度与顾客忠诚度之间的关系。 3.探讨人口统计变项与消费型态变项是否有相关。 4.探讨人口统计变项与消费型态变项跟顾客满意度是否有相关。 5.探讨人口统计变项与消费型态变项跟顾客忠诚度是否有相关。二、文献回顾连锁咖啡厅经营风格分类根据行政院（1996）所颁布的「中华民国行业标准分类」，咖啡厅是属於九大行业中的商业类之饮食业。而国内咖啡厅由於创业背景、风格以及产品组合等方面有其独特的特质，使得经营型态与风格呈现多元化的风貌。依照中华民国连锁店协会（1999）对咖啡产业调查指出，台湾目前的咖啡厅可分成以下几类：

2019年TPO20托福口语Task3原文及参考答案

2019年TPO20托福口语Task3原文及参考答案 TPO20口语Task3题目： Energy costs reach sky-high levels Heating prices, and the university’s heating bills, reached record-high levels this year. In an effort to reduce heating costs, the university plans to lower the temperature in all classroom buildings on weekends and on weekdays after five P.M., when the day’s classes have officially ended. Because classes will not be in session at these times, the reduced hear should have no negative impact on students and their studies. The university considers it essential that the cost-saving measures not interfere with student activities. ConversationNow listen to two students discussing the article. Woman： What on earth is the university thinking? Man： Huh? What do you mean? Woman：Well, those buildings aren’t just used for classes. That’s where my evening st udy group usually meets. Man：Oh, of course, that’s true. But you could go to the library instead. Woman： Yeah, but the library only has a few rooms that people can use for group meetings which is why classrooms are so convenient. But now, if it’s wint er and cold, where can we go?

高中外研社英语选修六Module5课文Frankenstein's Monster

Frankenstein's Monster Part 1 The story of Frankenstein Frankenstein is a young scientist/ from Geneva, in Switzerland. While studying at university, he discovers the secret of how to give life/ to lifeless matter. Using bones from dead bodies, he creates a creature/ that resembles a human being/ and gives it life. The creature, which is unusually large/ and strong, is extremely ugly, and terrifies all those/ who see it. However, the monster, who has learnt to speak, is intelligent/ and has human emotions. Lonely and unhappy, he begins to hate his creator, Frankenstein. When Frankenstein refuses to create a wife/ for him, the monster murders Frankenstein's brother, his best friend Clerval, and finally, Frankenstein's wife Elizabeth. The scientist chases the creature/ to the Arctic/ in order to destroy him, but he dies there. At the end of the story, the monster disappears into the ice/ and snow/ to end his own life. Part 2 Extract from Frankenstein It was on a cold November night/ that I saw my creation/ for the first time. Feeling very anxious, I prepared the equipment/ that would give life/ to the thing/ that lay at my feet. It was already one/ in the morning/ and the rain/ fell against the window. My candle was almost burnt out when, by its tiny light，I saw the yellow eye of the creature open. It breathed hard, and moved its arms and legs. How can I describe my emotions/ when I saw this happen? How can I describe the monster who I had worked/ so hard/ to create? I had tried to make him beautiful. Beautiful! He was the ugliest thing/ I had ever seen! You could see the veins/ beneath his yellow skin. His hair was black/ and his teeth were white. But these things contrasted horribly with his yellow eyes, his wrinkled yellow skin and black lips. I had worked/ for nearly two years/ with one aim only, to give life to a lifeless body. For this/ I had not slept, I had destroyed my health. I had wanted it more than anything/ in the world. But now/ I had finished, the beauty of the dream vanished, and horror and disgust/ filled my heart. Now/ my only thoughts were, "I wish I had not created this creature, I wish I was on the other side of the world, I wish I could disappear!” When he turned to look at me, I felt unable to stay in the same room as him. I rushed out, and /for a long time/ I walked up and down my bedroom. At last/ I threw myself on the bed/ in my clothes, trying to find a few moments of sleep. But although I slept, I had terrible dreams. I dreamt I saw my fiancée/ walking in the streets of our town. She looked well/ and happy/ but as I kissed her lips，they became pale, as if she were dead. Her face changed and I thought/ I held the body of my dead mother/ in my arms. I woke, shaking with fear. At that same moment，I saw the creature/ that I had created. He was standing/by my bed/ and watching me. His

人教版高中语文必修必背课文精编WORD版

人教版高中语文必修必背课文精编W O R D版 IBM system office room 【A0816H-A0912AAAHH-GX8Q8-GNTHHJ8】

必修1 沁园春·长沙（全文）毛泽东独立寒秋，湘江北去，橘子洲头。看万山红遍，层林尽染，漫江碧透，百舸争流。鹰击长空，鱼翔浅底，万类霜天竞自由。怅寥廓，问苍茫大地，谁主沉浮。携来百侣曾游，忆往昔峥嵘岁月稠。

恰同学少年，风华正茂，书生意气，挥斥方遒。指点江山，激扬文字，粪土当年万户侯。曾记否，到中流击水，浪遏飞舟。雨巷（全文）戴望舒撑着油纸伞，独自彷徨在悠长、悠长又寂寥的雨巷，我希望逢着一个丁香一样地

结着愁怨的姑娘。她是有丁香一样的颜色，丁香一样的芬芳，丁香一样的忧愁，在雨中哀怨，哀怨又彷徨；她彷徨在这寂寥的雨巷，撑着油纸伞像我一样，像我一样地默默彳亍着冷漠、凄清，又惆怅。她默默地走近，走近，又投出太息一般的眼光

她飘过像梦一般地，像梦一般地凄婉迷茫。像梦中飘过一枝丁香地，我身旁飘过这个女郎；她默默地远了，远了，到了颓圮的篱墙，走尽这雨巷。在雨的哀曲里，消了她的颜色，散了她的芬芳，消散了，甚至她的太息般的眼光丁香般的惆怅。撑着油纸伞，独自

彷徨在悠长、悠长又寂寥的雨巷，我希望飘过一个丁香一样地结着愁怨的姑娘。再别康桥（全文）徐志摩轻轻的我走了，正如我轻轻的来；我轻轻的招手，作别西天的云彩。那河畔的金柳，是夕阳中的新娘；波光里的艳影，在我的心头荡漾。软泥上的青荇，油油的在水底招摇；在康河的柔波里，我甘心做一条水草！那榆荫下的一潭，不是清泉，是天上虹揉碎在浮藻间，沉淀着彩虹似的梦。寻梦？撑一支长篙，向青草更青处漫溯，满载一船星辉，在星辉斑斓里放歌。

2019年TPO22托福口语Task3原文及参考答案

2019年TPO22托福口语Task3原文及参考答案 TPO22口语Task3题目： University Announces Change in Campus Tour Guides Traditionally, staff members from the admissions office have led the campus tours for secondary-school students who are considering attending the university. Beginningnext semester, however, current students will lead these guided walks around the campus instead. “The staff of the admissions office no longer have time to lead the campus tours. We need to focus our resources on our other responsibilities,” explain Christine Tanner, the director of admissions.“Furthermore, we feel that current students will offer unique and valuable insight to our visitors during these campus tours.” Conversation Now listen to two students discussing the article. Man： So, what do you think of this? Woman：Well, I think it’s a good idea, necessary, too. Man： Why is that? Woman：Well, you know, I’ve actually been up to the admissions office a couple of times, like, to pick up some information for a friend, you know, and every time I’ve ever been there, wow, you can tell they’re really busy. Man： Oh, really.

(完整版)Unit7TheMonster课文翻译综合教程四

Unit 7 The Monster Deems Taylor 1He was an undersized little man, with a head too big for his body ― a sickly little man. His nerves were bad. He had skin trouble. It was agony for him to wear anything next to his skin coarser than silk. And he had delusions of grandeur. 2He was a monster of conceit. Never for one minute did he look at the world or at people, except in relation to himself. He believed himself to be one of the greatest dramatists in the world, one of the greatest thinkers, and one of the greatest composers. To hear him talk, he was Shakespeare, and Beethoven, and Plato, rolled into one. He was one of the most exhausting conversationalists that ever lived. Sometimes he was brilliant; sometimes he was maddeningly tiresome. But whether he was being brilliant or dull, he had one sole topic of conversation: himself. What he thought and what he did. 3He had a mania for being in the right. The slightest hint of disagreement, from anyone, on the most trivial point, was enough to set him off on a harangue that might last for hours, in which he proved himself right in so many ways, and with such exhausting volubility, that in the end his hearer, stunned and deafened, would agree with him, for the sake of peace. 4It never occurred to him that he and his doing were not of the most intense and fascinating interest to anyone with whom he came in contact. He had theories about almost any subject under the sun, including vegetarianism, the drama, politics, and music; and in support of these theories he wrote pamphlets, letters, books ... thousands upon thousands of words, hundreds and hundreds of pages. He not only wrote these things, and published them ― usually at somebody else’s expense ― but he would sit and read them aloud, for hours, to his friends, and his family. 5He had the emotional stability of a six-year-old child. When he felt out of sorts, he would rave and stamp, or sink into suicidal gloom and talk darkly of going to the East to end his days as a Buddhist monk. Ten minutes later, when something pleased him he would rush out of doors and run around the garden, or jump up and down off the sofa, or stand on his head. He could be grief-stricken over the death of a pet dog, and could be callous and heartless to a degree that would have made a Roman emperor shudder. 6He was almost innocent of any sense of responsibility. He was convinced that

人教版高中语文必修一背诵篇目

高中语文必修一背诵篇目 1、《沁园春长沙》毛泽东独立寒秋，湘江北去，橘子洲头。看万山红遍，层林尽染；漫江碧透，百舸争流。鹰击长空，鱼翔浅底，万类霜天竞自由。怅寥廓，问苍茫大地，谁主沉浮？携来百侣曾游，忆往昔峥嵘岁月稠。恰同学少年，风华正茂；书生意气，挥斥方遒。指点江山，激扬文字，粪土当年万户侯。曾记否，到中流击水，浪遏飞舟？ 2、《诗两首》（1）、《雨巷》戴望舒撑着油纸伞，独自 /彷徨在悠长、悠长/又寂寥的雨巷，我希望逢着 /一个丁香一样的 /结着愁怨的姑娘。她是有 /丁香一样的颜色，/丁香一样的芬芳， /丁香一样的忧愁，在雨中哀怨， /哀怨又彷徨； /她彷徨在这寂寥的雨巷，撑着油纸伞 /像我一样， /像我一样地 /默默彳亍着冷漠、凄清，又惆怅。 /她静默地走近/走近，又投出太息一般的眼光，/她飘过 /像梦一般地， /像梦一般地凄婉迷茫。像梦中飘过 /一枝丁香的， /我身旁飘过这女郎；她静默地远了，远了，/到了颓圮的篱墙， /走尽这雨巷。在雨的哀曲里， /消了她的颜色， /散了她的芬芳， /消散了，甚至她的太息般的眼光， /丁香般的惆怅/撑着油纸伞，独自 /彷徨在悠长，悠长又寂寥的雨巷， /我希望飘过 /一个丁香一样的 /结着愁怨的姑娘。（2）、《再别康桥》徐志摩轻轻的我走了， /正如我轻轻的来； /我轻轻的招手， /作别西天的云彩。那河畔的金柳， /是夕阳中的新娘； /波光里的艳影， /在我的心头荡漾。软泥上的青荇， /油油的在水底招摇； /在康河的柔波里， /我甘心做一条水草！那榆阴下的一潭， /不是清泉，是天上虹 /揉碎在浮藻间， /沉淀着彩虹似的梦。寻梦？撑一支长篙， /向青草更青处漫溯， /满载一船星辉， /在星辉斑斓里放歌。但我不能放歌， /悄悄是别离的笙箫； /夏虫也为我沉默， / 沉默是今晚的康桥！悄悄的我走了， /正如我悄悄的来；/我挥一挥衣袖， /不带走一片云彩。 4、《荆轲刺秦王》太子及宾客知其事者，皆白衣冠以送之。至易水上，既祖，取道。高渐离击筑，荆轲和而歌，为变徵之声，士皆垂泪涕泣。又前而为歌曰：“风萧萧兮易水寒，壮士一去兮不复还！”复为慷慨羽声，士皆瞋目，发尽上指冠。于是荆轲遂就车而去，终已不顾。 5、《记念刘和珍君》鲁迅（1）、真的猛士，敢于直面惨淡的人生，敢于正视淋漓的鲜血。这是怎样的哀痛者和幸福者？然而造化又常常为庸人设计，以时间的流驶，来洗涤旧迹，仅使留下淡红的血色和微漠的悲哀。在这淡红的血色和微漠的悲哀中，又给人暂得偷生，维持着这似人非人的世界。我不知道这样的世界何时是一个尽头！

星巴克swot分析

6月21日 85度C VS. 星巴克SWOT分析星巴克SWOT分析优势 1.人才流失率低 2.品牌知名度高 3.熟客券的发行 4.产品多样化 5.直营贩售 6.结合周边产品 7.策略联盟劣势 1.店內座位不足 2.分店分布不均机会 1.生活水准提高 2.隐藏极大商机 3.第三空间的概念 4.建立电子商务威胁 1.WTO开放后，陆续有国际品牌进驻 2.传统面包复合式、连锁咖啡馆的经营星巴克五力分析１、供应商：休闲风气盛，厂商可将咖啡豆直接批给在家煮咖啡的消费者２、购买者：消费者意识高涨、资讯透明化（比价方便）３、同业內部竞争：产品严重抄袭、分店附近必有其他竞争者４、潜在竞争者：设立咖啡店连锁店无进入障碍、品质渐佳的铝箔包装咖啡５、替代品：中国茶点、台湾小吃、窜红甚快的日本东洋风...等８５度ｃ市場ｓｗｏｔ分析：【Strength优势】具合作同业优势产品精致以价格进行市场区分，平价超值服务、科技、产品、行销创新，机动性强加盟管理人性化【Weakness弱势】通路品质控制不易品牌偏好度不足通路不广财务能力不健全 85度c的历史资料，在他们的网页上的活动信息左邊的新聞訊息內皆有詳細資料，您可以直接上網站去查閱，皆詳述的非常清楚。

顧客滿意度形成品牌重於產品的行銷模式。你可以在上他們家網站找找看！【Opportunity机会】勇于變革变革与创新的经营理念同业策略联盟的发展、弹性空间大【Threat威胁】同业竞争对手(怡客、维多伦)门市面对面竞争加盟店水准不一，品牌形象建立不易直，间接竞争者不断崛起(壹咖啡、City Café．．．) 85度跟星巴克是不太相同的零售，星巴客应该比较接近丹堤的咖啡厅，策略形成的部份，可以从 1.产品线的宽度跟特色 2.市场区域与选择 3.垂直整合 4.规模经济 5.地区 6.竞争优势這6點來做星巴客跟85的区分可以化一个表，來比较他們有什么不同內外部的話，內部就从相同产业來分析(有什麼优势跟劣势) 外部的話，一樣是相同产业但卖的东西跟服务不太同，来与其中一个产业做比较(例如星巴客) S(优势):點心精致平价,咖啡便宜,店面设计观感很好...等等 W(劣势):对于消费能力较低的地区点心价格仍然较高,服务人员素质不齐,點心种类变化較少 O(机会):对于点心&咖啡市场仍然只有少数的品牌独占(如:星XX,壹XX...等),它可以透过连锁店的开幕达成高市占率 T(威协):1.台湾人的模仿功力一流,所以必须保持自己的特色做好市场定位 2.消費者的口味变化快速,所以可以借助学者"麥XX"的做法保有主要的點心款式外加上几样周期变化的點心五力分析客戶讲价能力（the bargaining power of customers）、供应商讲价能力（the bargaining power of suppliers）、新进入者的竞争（the threat of new entrants）、替代品的威协（the threat of substitute products）、现有厂商的竞争（The intensity of competitive rivalry）

托福综合口语TASK2各类实用模板分享

托福综合口语TASK2各类实用模板分享托福口语和写作一样，都是属于需要考生自己组织语言来进行表达输出的考试题型。今天给大家带来了托福综合口语TASK2各类实用模板分享,希望能够帮助到大家，下面就和大家分享，来欣赏一下吧。【精选素材】托福综合口语TASK2各类实用模板分享托福综合口语TASK2备考模版 1、“The school has implemented a new policy that ... due to ... . And the man/woman holds a positive/negative view towards the announcement. The first reason s/he gives is that ... . And the second one is based on the fact that ... .” 2、From the reading material, we know that (the college) is going to ... Obviously, the man/woman in the conversation thinks that this is a great/bad idea, due to the following reasons..

One reason is that……. Another is,……. (万一有时间)So thats all the reasons s/he has to form that opinion. (尽量记，听为主) 3、The man/woman’s opinion about STH is that……. (选择方案型：听选择的原因。1.Firstly, the man/woman states that…….2.And then he/she states that…….) (提出建议型：听建议的利弊。1.Firstly, the man/woman talks about the positive aspect of this issue. He/she states that…….2.On the other hand, the man/man also talks about the negative aspect of this question. He/she points out that…….) 4、阅读引用部分的模版 (1) In the readingmaterial, (2) There is a/anannouncement/message/notice/proposal about ..记下的关键词) (3) The university/college is going to..(稍稍展开下) (4) In the listeningmaterial, two students discussabout the..填入关键词) 听力中人物表达观点的模版:

人教版高中语文必修必背课文

必修1 沁园春·长沙（全文）毛泽东独立寒秋，湘江北去，橘子洲头。看万山红遍，层林尽染，漫江碧透，百舸争流。鹰击长空，鱼翔浅底，万类霜天竞自由。怅寥廓，问苍茫大地，谁主沉浮。携来百侣曾游，忆往昔峥嵘岁月稠。恰同学少年，风华正茂，书生意气，挥斥方遒。指点江山，激扬文字，粪土当年万户侯。曾记否，到中流击水，浪遏飞舟。雨巷（全文）戴望舒撑着油纸伞，独自彷徨在悠长、悠长又寂寥的雨巷，我希望逢着一个丁香一样地结着愁怨的姑娘。她是有丁香一样的颜色，丁香一样的芬芳，丁香一样的忧愁，在雨中哀怨，哀怨又彷徨；

她彷徨在这寂寥的雨巷，撑着油纸伞像我一样，像我一样地默默彳亍着冷漠、凄清，又惆怅。她默默地走近，走近，又投出太息一般的眼光她飘过像梦一般地，像梦一般地凄婉迷茫。像梦中飘过一枝丁香地，我身旁飘过这个女郎；她默默地远了，远了，到了颓圮的篱墙，走尽这雨巷。在雨的哀曲里，消了她的颜色，散了她的芬芳，消散了，甚至她的太息般的眼光丁香般的惆怅。撑着油纸伞，独自彷徨在悠长、悠长又寂寥的雨巷，我希望飘过一个丁香一样地结着愁怨的姑娘。再别康桥（全文）徐志摩轻轻的我走了，正如我轻轻的来；我轻轻的招手，作别西天的云彩。那河畔的金柳，是夕阳中的新娘；波光里的艳影，在我的心头荡漾。软泥上的青荇，油油的在水底招摇；

在康河的柔波里，我甘心做一条水草！那榆荫下的一潭，不是清泉，是天上虹揉碎在浮藻间，沉淀着彩虹似的梦。寻梦？撑一支长篙，向青草更青处漫溯，满载一船星辉，在星辉斑斓里放歌。但我不能放歌，悄悄是别离的笙箫；夏虫也为我沉默，沉默是今晚的康桥。悄悄的我走了，正如我悄悄的来；我挥一挥衣袖，不带走一片云彩。记念刘和珍君（二、四节）鲁迅二真的猛士，敢于直面惨淡的人生，敢于正视淋漓的鲜血。这是怎样的哀痛者和幸福者？然而造化又常常为庸人设计，以时间的流驶，来洗涤旧迹，仅使留下淡红的血色和微漠的悲哀。在这淡红的血色和微漠的悲哀中，又给人暂得偷生，维持着这似人非人的世界。我不知道这样的世界何时是一个尽头！我们还在这样的世上活着；我也早觉得有写一点东西的必要了。离三月十八日也已有两星期，忘却的救主快要降临了罢，我正有写一点东西的必要了。四我在十八日早晨，才知道上午有群众向执政府请愿的事；下午便得到噩耗，说卫队居然开枪，死伤至数百人，而刘和珍君即在遇害者之列。但我对于这些传说，竟至于颇为怀疑。我向来是不惮以最坏的恶意，来推测中国人的，然而我还不料，也不信竟会下劣凶残到这地步。况且始终微笑着的和蔼的刘和珍君，更何至于无端在府门前喋血呢？然而即日证明是事实了，作证的便是她自己的尸骸。还有一具，是杨德群君的。而且又证明着这不但是杀害，简直是虐杀，因为身体上还有棍棒的伤痕。但段政府就有令，说她们是“暴徒”！但接着就有流言，说她们是受人利用的。惨象，已使我目不忍视了；流言，尤使我耳不忍闻。我还有什么话可说呢？我懂得衰亡民族之所以默无声息的缘由了。沉默啊，沉默啊！不在沉默中爆发，就在沉默中灭亡。