文档库 最新最全的文档下载
当前位置:文档库 › CBR Approach

CBR Approach

CBR Approach
CBR Approach

Technical Report IIIA-2000-04. Extended version of a paper to be published in the special issue on prognosis of "Methods of Information in Medicine"

Individual prognosis of diabetes long-term risks: A

CBR Approach

E. Armengol1, A. Palaudàries2 and E. Plaza1*

1Artificial Intelligence Research Institute (IIIA-CS IC)

Campus UAB, 08193- Bellaterra, Catalonia, S pain

2 Unitat d'Endocrinologia. Hospital de Mataró

Carretera de Cirera, s/n. 08304-Mataró, Catalonia, Spain

Abstract

In this paper we present DIRAS, an application supporting the physicians to determine the risk of complications for individual diabetic patients. The risk pattern of each diabetic patient is ob-tained using a Case-based Reasoning method called LID. Case-based Reasoning is an Artificial In-telligence technique based on solving new situations according to past experiences. For each patient, the LID method determines the risk of each diabetic complication according to the risk of already diagnosed patients. In addition, LID builds a description that can be viewed as an explanation of the obtained risk.

1.Introduction

Lucas and Abu-Hanna [1] define prognosis as the prediction of the course and outcome of disease processes. Usually, computer systems supporting the physicians to take decisions use prognosis mod-els that predict the outcomes of a particular process. Techniques used to build prognosis models range from a construction by hand to statistical techniques or Artificial Intelligence (AI) techniques –see [1] for the description of specific techniques. In particular, we have been investigating Case-based Reasoning [2] an AI technique based in the human capability to solve new situations by learn-ing from the past situations already solved.

In this paper we present DIRAS (Diabetes Individualized Risk Assessment System), an ap-plication whose goal is to predict the risk of complications for diabetic patients. Diabetes Mellitus

*Corresponding author. Phone: 34–935809570, Fax: 34 - 935809661,e-mail: enric@iiia.csic.es

is one of the most frequent human clinic diseases since it affects around a 3% of the European popu-lation and around one hundred million people in the world. There are two major types of diabetes: type 1 (or insulin-dependent) and type 2 (or non insulin-dependent). The diabetes type 1 usually de-velops in children or people less than 40 years old. This form of diabetes is characterized by an in-sufficient production of insulin at the pancreas. People with this type of diabetes need daily injec-tions of insulin. If not diagnosed and treated with insulin, the person can lapse into a life-threatening coma. Diabetes type 2, the most common one, usually develops in adults over the age of 40 being more common among adults over 55. Usually, people with diabetes type 2 have overweight and sedentary lifestyle. In diabetes type 2 the pancreas produces insulin but the body does not uses it effectively. The consequences are the same that those of the diabetes type 1 although its symp-toms appear gradually, and they tend to be vague. Some people with diabetes type 2 must inject in-sulin, but most are controlled with a combination of weight loss, exercise, and prescription of oral diabetes medication.

A bad management of both forms of diabetes will produce microcomplications (such as blindness, renal failure or polyneuropathy), and macrocomplications (such as gangrene and amputa-tion, aggravated coronary heart disease or stroke). Therefore, main concern in the management of the diabetes is reducing the risk of a patient developing a new long-term complication and the risk of progression in the complications already present. The prediction of the individual risk to de-velop long-term complications is based on the analysis of a large quantity of data (e.g. diabetes type, diabetes duration, age, cholesterol, and metabolic control degree) that have to be continu-ously evaluated. The therapeutic goals to offer a good life quality to the patient depend on this analysis. Because diabetes mellitus has a high prevalence, sometimes physicians taking care of diabetic patients have not a specialized formation in diabetes and, consequently, the management of these patients may be less accurate. For this reason, a system supporting an individualized as-sessment of the patients can be a useful tool to improve both the management and the treatment of diabetes.

We have developed the DIRAS application supporting the physicians to determine the risk of complications for each patient according to the clinical data of that patient. We call risk pattern to the set of assessments concerning diabetic complications. DIRAS obtains a risk pattern where the risk of diabetic macrocomplications (ischemic cardiopathy, low extremities vasculopa-thy, and stroke) and diabetic microcomplications (nephropathy, retinopathy and polyneuropathy) are explicitly assessed. The DiabCare Quality Network (http://www.diabcare.de) is a European consortium having as goal the improvement of the care in diabetes type 1 and type 2. Basically, the goal of this project is to implement effective measures for the prevention of complications such as blindness due to diabetes, number of people entering to an end-stage diabetic renal failure, etc. DiabCare manages groups of diabetic patients using statistical tools. The main contribution of DIRAS is focusing on individual patients instead of populations of patients. For this purpose, we

use an Artificial Intelligence technique called Case-based Reasoning(2) for assessing the risk of complications in patients with diabetes type 1 or type 2.

The structure of this paper is the following. Section 2 explains the goal and the knowledge structures used by the DIRAS application. Section 3 presents LID, a Case-based Reasoning method used by DIRAS to assess the complication risks of a patient. Finally, in section 4 the results of DIRAS are discussed.

2. The DIRAS Application

The goal of DIRAS is to determine the risk of complications for individual diabetic patients, what we will call the risk pattern. For each patient, DIRAS works with five kinds of data (Figure 1): Personal-Data, Basic-Diabetes-Data, Info-Patient-Consultation, Assessment and Risk-Pattern. Personal-data contains information such as the name, address, birth date, etc. Basic-Diabetes-Data contains basic information of diabetes (such as diabetes type, duration, and whether diabetes is treated with oral drugs or insulin). Info-Patient-Consultation has data on relevant measures (e.g. glycated hemoglobin, cholesterol, blood pressure, etc), eye and foot examination, current treatments, etc.

Figure 1. Description of a patient

Assessment contains knowledge obtained by applying domain knowledge provided by an expert dia-betologist. This domain knowledge allows the analysis of the patient's data obtaining a high level perspective of the patient's state. For instance, Assessment holds a qualitative measure of the LDL-cholesterol (q-LDL-chol of Assessment in Fig. 1). This qualitative measure may have a value low, moderate or high depending on the following conditions:

if the patient has macrocomplications then

if LDL-chol > 130 then q-LDL-chol = high

else if LDL-chol > 100 then q-LDL-chol = moderate

else low

if the patient has no macrocomplications then

if LDL-chol > 150 then q-LDL-chol = high

else if LDL-chol > 130 then q-LDL-chol = moderate

else low

Notice that the qualitative measure depends on the presence or absence of macrocomplications. DIRAS has similar rules for qualifying other measures in an appropriate way.

In addition, from some measures and symptoms, DIRAS can infer for Assessment new facts like whether or not the patient has a specific complication. For instance the feature micro-compl? of Assessment contains information about the presence or absence of microcomplications in the current patient. This feature has value true if the patient has eye lesions, nephropathy or polyneuropa-thy; otherwise the feature micro-compl? has value false. In turn, the presence or absence of eye le-sions, nephropathy or polyneuropathy is inferred by DIRAS using similar rules that take into ac-count relevant data held in Info-Patient-Consultation.

Finally, the last kind of patient data is Risk-Pattern i.e. the assessment of individual long-term risks that we want to estimate using Case-based Reasoning. Risk-Pattern has two parts (Figure 1): 1) the macrocomplication risks, and 2) the microcomplication risks. For macrocomplications we want to assess both the global risk and the risk of three particular macrocomplications (namely stroke, infarct and amputation). S imilarly, for microcomplications we want to assess both the global risk and the risk of polyneuropathy, nephropathy and retinopathy. The global risk repre-sents the risk of a patient to vascular alterations that have not defined symptoms such as the intes-tine infarct.

There are two kinds of risk for complications: development risk and progression risk. The development risk has to do with patient's likelihood of developing a new complication in the fu-ture. The progression risk is when a patient already has a macrocomplication and thus the risk of further deterioration has to be assessed.

The next section shows in detail how DIRAS uses Case-based Reasoning to obtain a risk pat-tern for individual patients.

3. CBR Assessment Risk Pattern

The goal of DIRAS is to obtain an individual risk pattern for diabetic patients using Case-based Reasoning [2]. Case-based Reasoning (CBR) is an AI technique based on the human capability to solve new situations according to past experience. The core idea of CBR is that when a new situation is similar to one or several old situations, the decisions taken and the knowledge contained in old situations provide a starting point to interpret or solve the new situation. Each situation is called a case (or precedent) that may be reused to solve new problems. The collection of cases of a system is called the case base. The description of a new situation to be solved is called the current case or problem.

Given a case base and a problem, CBR methods perform three tasks [3]: 1) retrieve, that ob-tains past cases similar to the new case; 2) select, that decides which of the retrieved past cases is the most similar (i.e. the best precedent) to the current problem; and 3) adapt, that decides how to adapt the solution of the best precedent to solve the current problem.

DIRAS uses a case base where each case is a patient described as explained in the previous section, i.e. the patient's data plus the solution (the risk pattern) for that patient. The goal of DIRAS is to obtain the risk pattern for the current patient. S everal features (see Figure 1) form the risk pattern and DIRAS obtains the risk for each feature in an independent way.

Task Method

Figure 2. Task decomposition of the risk-assessment task that obtains the risk pattern of diabetic

patients.

The complete risk pattern is obtained by solving the Risk-Assessment task that decomposes in two tasks (Fig. 2): the Macro-Risk-Assessment task and the Micro-Risk-Assessment task. The Macro-Risk-Assessment task decomposes, in turn, in three subtasks: the Kind task determines whether the risk of macrocomplications is progression or development; the Global-Macro task assesses the global risk of macrocomplications; and the Specific-Risks task assesses the risk of three specific macrocomplica-tions, namely infarct, stroke and amputation. S imilarly, the Micro-Risk-Assessment task is decom-posed in three subtasks: Kind,Global-Micro and Specific-Risks. In particular the Specific-Risks task as-sesses the risk of three specific microcomplications: retinopathy, polyneuropathy and nephropa-thy. All these risks are inferred using LID (Lazy Induction of Descriptions), a Case-based Reasoning method that is explained in the following section.

3.1. The LID Method

In this section we introduce LID, the CBR method used by DIRAS to solve the risk assessment tasks described in Fig. 2. For each diabetic task LID searches the case base for the best precedent and in-fers the risk according to that precedent.

For a given collection of risk classes R = {unknown, low, moderate, high, very-high}, a dia-betic complication C, and a problem p, the task of LID is to obtain the risk R i ∈ R of p concerning C. For each complication C, this can be seen as a classification task where the goal is to identify the class in R to which p belongs. DIRAS solves this classification task using LID.

Given a case base B containing diabetic patients classified into the collection of risk classes R for a diabetic complication C, and a problem p to be classified, LID obtains the class R i ∈R to which p belongs. Intuitively, LID follows a top-down strategy to build a description D containing the most relevant features of p such that all features in D are satisfied by a subset of cases in B. In general, cases in this subset belong to different solution classes in R. LID adds relevant features to D until the subset of cases satisfying D belong to one unique solution class R i. LID takes this class R i as the solution for the current task, i.e. R i is the risk of p concerning C.

D:= ?; R = {R

1… R

n

}

Function LID (B, p, D)

S

D

:= Discriminatory-set (D, B)

if?e

i

∈S D ?e i∈R i then return R i

else f

d

:= Select-Feature (p, B, R)

D':= Add-Feature (f

d

, D)

LID (S

D

, p, D')

end-if

end-function

Figure 3. The LID algorithm

The LID algorithm (Fig. 3) begins with the whole set of precedents B classified into the collection of risk classes R for a complication C, a problem p to be solved and the description D = ? (i.e. D has no features). In the following we will explain this algorithm using an example. See [4] for more de-tailed explanation and some results of the LID method on other domains of application.

Example 1. Let p be a patient with no macrocomplications (i.e. feature macro-compl? in Assessment has value false), high blood pressure and low albumin. In this example DIRAS has to determine the risk R i ∈ R for the macrocomplication C = stroke.

The set of cases S D ? B that are subsumed by the description D is called discriminatory set. Intuitively, a case c is subsumed by a description D when all the information contained in D is also contained in c, although c can contain more information. See a formal description of subsumption in [5].

Initially D is an empty description, i.e. it is the most general description. Therefore D sub-sumes all the cases in B (i.e. S D = B), and consequently D has to be specialized. The specialization of a description D is achieved by adding features to it. In particular, LID adds a feature f with the value v that this feature has in the current problem p. After that, the new description D' = D + (f=v) has a smaller discriminatory set S D' formed by those cases subsumed by D'. Thus, specializa-tion reduces the discriminatory set S D'? S D at each step. As we explain below, LID uses a heuristic measure based on the López de Mántaras distance [6] to determine the feature to be added.

LID specializes D by selecting one feature f from all the features used in p in the following way. Each feature f i in p induces a partition P i = {S i1… S in} in the set S D such that each S ik ∈ P i con-tains those patients in S D having the same value v k in the feature f i. For instance, the presence or absence of macrocomplications will divide the set S D (currently S D = B) in two subsets: one contain-ing those precedents having macrocomplications and the other one containing patients without macrocomplications. There is also a partition of S D, called the correct partition P c, that divides S D according to the risk (R i ∈ R) for the complication C. In the example, S D is divided in subsets accord-ing to the values for the risk of stroke being unknown, low, moderate, high, and very-high.

For each partition P i, LID computes the López de Mántaras (RLM) distance [6] to the correct partition P c. Intuitively, the RLM distance assesses how similar a partition is with respect to a ref-erent partition (i.e. the correct partition), in the sense that the lesser the distance the more similar they are. The RLM distance was introduced as an alternative to the Quinlan’s Gain [7] used in the ID3 inductive learning algorithm. The Quinlan’s Gain is a selection measure that selects the object feature providing the highest information gain. RLM distance shows that normalizing the Quin-lan's Gain in an appropriate way, we obtain a distance between partitions.

Formally, given two partitions P i and P c of a set S D, the RLM distance between them is com-puted as follows:

RLM(,P )()()()c P I P I P I P P where I P p p p S S S I P p p p S R S I P P p p i i c i c i j j j n j D ij D c k k k m k D k D

i c jk =?+∩=?=∩=?=

∩∩=?==∑∑221

212 ()log ;()log ,()log jk k m j n jk D k ij D

p S R S S ==∑∑

=

∩∩11 where I(P i ) measures the information contained in the partition P i ; n is the number of possible values of the feature inducing P i ; m = Card(R); p i is the probability of occurrence of class S ij (R j ) i.e. the proportion of examples in S D that belong to S ij (R j ); I(P i ∩ P c ) is the mutual information of two parti-tions; and p jk is the probability of occurrence of the intersection R j ∩ S ik , i.e. the proportion of ex-amples in S D that belong to R j and to S ik .

Using the RLM distance, we can define what it means for a feature to be more discriminatory than another.

Definition ["More discriminatory than" relation]. Let P c be the correct partition (i.e. the partition

that correctly classifies the examples), and P j and P k the partitions induced by features f j and f k respectively, we say that feature f j is more discriminatory than feature f k iff RLM (P j , P c )

In other words, when a feature f j is more discriminatory than another feature f k the partition that f j induces in S D is closer to the correct partition P c than the partition induced by f k . Intuitively, the most discriminatory feature classifies the cases in S D in a more similar way to the correct classifica-tion of cases (i.e. that determined by the risk of the complication C).

Thus, LID selects the feature f having minimum RLM distance to the correct partition as be-ing the most discriminatory. Then, LID builds D', a specialization of D, by adding to D the feature f with the same value that f takes in the current problem p . In the example, the most discriminatory feature is macro-compl? Thus the description D' will contain the feature macro-compl? with value false (since the current problem p corresponds to a patient that has no macrocomplications).

Let S D' be the subset of precedents that subsumed by D'. If all the precedents in S D' belong to only one risk class R i then LID finishes the process and classifies p as belonging to R i . Otherwise, D'needs to be further specialized. In Example 1, S D' contains those cases in S D = B with the feature macro-compl? being false . However, with respect to the correct partition P c (induced from the risk of stroke) the cases in S D' belong to several of these solution classes in R. Because of this, D' need be specialized in order to reduce the discriminatory set S D'. This specialization is made using LID with the set S D', the description D' and the patient p . Notice that LID may safely ignore the precedents

that do not belong to S D' because the precedents that are not subsumed by D' will not be subsumed by any specialization of D'.

The next step in the specialization of D' is the selection of the next most discriminatory fea-ture. Now LID finds that the most discriminatory feature is the qualitative measure of the blood pressure (feature q-bp of Assessment in Figure 1). The patient p has value high in the q-bp feature, therefore the specialization D'' of D' contains two features: macro-compl? with value false (as be-fore) and the newly added feature q-bp with value high.

Subsequently, LID considers the discriminatory set S D'' containing those precedents in S D' sub-sumed by D'' and finds whether or not all cases in S D'' belong only to one class R i. In Example 1, a l l the cases contained in S D'' belong to the class R i = high for the risk of stroke. Therefore, LID finishes inferring that the patient p has also a high risk of stroke.

There is an abnormal stopping condition produced when there is no possible to specialize the current description. This situation occurs when LID has used all the features candidates to spe-cialize a description but the current description D n subsumes precedents belonging to several classes R' ? R. In that situation, LID proposes as solution for the current problem the classes in R'.

Figure 4. Browsing of a Risk-Pattern obtained by DIRAS in the example 1.

DIRAS uses LID to solve the tasks shown in Figure . That is to say, for macrocomplications LID is used to solve the tasks Kind, Global-Macro, Infarct, Stroke, and Amputation; and for microcomplications LID is used to solve the tasks Kind, Global-Micro, Retinopathy, Polyneuropathy and Nephropathy. There-fore, for each diabetic patient, DIRAS obtains a risk pattern as the one in Fig. 4 for example 1.

Another concern of LID is the interpretation of the descriptions as an explanation for the assessment of a specific risk. For instance, Fig. 5 shows the description obtained by LID for Example 1. This description considers that macro-compl? and q-bp are relevant to assess the risk of stroke. In particular, the expert diabetologist agrees with this explanation since it is consistent with his knowledge. For other examples, LID explains the moderate risk of stroke of a patient with no mac-rocomplications because of the moderate blood pressure. Nevertheless, when the patient has mac-

rocomplications and his blood pressure is moderate, LID assesses a high risk of stroke. The expert agrees with both explanations, since it is known that the risk of stroke directly depends on the level of the blood pressure and it is increased when the patient has macrocomplications. There also are examples to which DIRAS has assessed a high global progression risk of macrocomplications because the patient has high levels of both LDL-cholesterol and HDL-cholesterol. Moreover, DIRAS has obtained the same explanation to justify the high global progression risk of other cases. The expert agrees this explanation for these specific cases.

The estimation of pure accuracy allows us to distinguish the parts of DIRAS that require revisions (like the one related to infarct, that probably requires an enlargement of the case base) but it’s not a good way to validate a medical recommendation system. The final validation will follow the process and use the criteria developed at our Institute for validating the PneumonIA ex-pert system [8]. This process involves several expert diabetologists to which the data of the pa-tients in the case base are shown. Each expert independently assesses the risk pattern of each pa-tient, and in this way we find out areas of consensus and areas where risk assessment may have some admissible variation. Finally, the goal of this validation process is to establish whether the performance of DIRAS is indistinguishable or not from that of the human experts. S pecifically, we will perform a ranking of the set of composed of the experts plus DIRAS. If the system is ranked among the experts this means that its performance is indistinguishable from that of the best ex-perts, while if it ranks below them it is distinguishable [8].

3.3. Report of a Diabetic Patient

In addition to the risk pattern of an individual patient, DIRAS produces a report that can be useful for nonexpert physicians to manage diabetic patients. This report (Fig. 6) is formed by four sections: 1. personal data of the patient (e.g. age, diabetes type, year of the diabetes diagnostic), 2. as-sessments about the measures, 3. Information about macrocomplications, and 4. Information about microcomplications. Section 2 of the report contains, for each measure M, the patient's value for M, the range of normal values for M, an assessment on whether the patient's value for M is acceptable or not. Section 3 of the report is divided in two parts. The first part contains the report of the pa-tient's macrocomplications detailing some aspects of the patient's state (for instance the foot state). The second part shows the risk of each macrocomplication and the factors used to determine this risk. The section 4 of the report has the same structure that the section 3 but has to to do with the patient's state and risks concerning microcomplications.

___________________________________________________________________________________ 1- PATIENT DATA

Number: 3408301200246854

Man, 76 years old, with a diabetes type 2 diagnosed 25 years ago

___________________________________________________________________________________ 2- GENERAL ASSESSMENT

Value Good Value Assessment

HbA1c9.10<= 6.5 unacceptable

BP (SYS) 190.00<= 130 unacceptable

BP (DIA) 90.00<= 80 acceptable

Cholesterol 166.00<= 180 correct

LDL Chol__ <= 100 (macro)/ <= 130 __

HDL Chol__ => 45 __

Triglicerids 94.00<= 150correct

Albumin 289.00<= 20 unacceptable

Creatinine 101.00<= 106 correct

Good stability because he/she had no hypoglycemies nor hyperglycemies. No hospitalizations.

___________________________________________________________________________________ 3- MACROVASCULAR COMPLICATIONS

Ischemic heart : No

Infarct (or coronary bypass or angioplasty): No

Anginal chest pain : No

Stroke : No

Low extremities vasculopathy : Yes

Amputation above ankle : No

Amputation below ankle : No

Leg claudication : No Right foot Left foot

Bypass or angioplasty :No No

Feet pulse present :No No

Healed ulcer :No No

Acute ulcer :No No

___________________________________________________________________________________

Global risk progression: HIGH because the value of total cholesterol is high

___________________________________________________________________________________ Risk for Specific Macrocomplications :

Stroke : VERY-HIGH because the blood pressure is very high and the patient has macrocomplications

Infarct : HIGH because the value of total cholesterol is high

Lesion/amp. : HIGH, because the patient has polyneuropathy with normal sensitivity and vasculopathy

___________________________________________________________________________________ 4- MICROVASCULAR COMPLICATIONS

Polyneuropathy : Yes

Neuropathy symptoms : true Right foot Left foot

Pulse present : No No

Pin prich sensitivity :Abnormal Abnormal

Vibration sensitivity : Abnormal Abnormal

Nephropathy : Yes

Phase I: heavy MAU (MAU : 289)

Renal insuficiency : No (creatinine : 101)

Retinopathy : Yes Right Eye Left Eye

Retinopathy type : PREPROLIFERATIVE PREPROLIFERATIVE

Visual Acuity :Unknown Unknown

___________________________________________________________________________________ Risk for specific microcomplications :

Polyneuropathy : HIGH progression risk because the HbA1c is high

Nephropathy : VERY HIGH progression risk because the albumin is correct, the HbA1c is high, the blood pressure is not

low and the patient does not follow nephropathy treatment

Retinopathy : HIGH progression risk because the HbA1c is high

Visual Ac. Dec. : MODERATE progression risk because the maculopathy has been photocoagulated

Figure 6. Report of complications for a diabetic patient.

4. Related Work and Conclusions

There are two aspects of DIRAS that can be compared with other works: the application domain and the methodology. Concerning to the domain application, there are several applications used in the management of diabetic patients [9, 10]. These applications are oriented to determine the insu-lin dosage for a diabetic patient of type 1. Basically the goal is to determine a management plan for each patient according to both his particular lifestyle (i.e. diet and physical exercise) and his metabolic state (i.e. glucose levels). Instead, the goal of DIRAS is to assess the risk of long term complications for individual diabetic patients (either type 1 or type 2 diabetes).

Concerning the methodology, DIRAS uses LID, a Case-based Reasoning method that builds a discriminant explanation of the assessed risk of complication using an heuristic based on the RLM distance. This heuristic has been used in induction of decision trees [7]. The only similitude here is in the use of a heuristic for selecting an attribute as more discriminant than others, but the structure that is built in decision trees and in LID are different. Decision trees build a structure that classifies a training set of examples and uses the heuristic to select the branching criteria of that tree struc-ture. The case-based method LID is a problem-centered technique that builds a structure that dis-criminates the new problem with respect to classes in the training set and the heuristic is used to decide the attribute that is more useful in discriminating this new example with respect to the training examples.

One of the advantages of Case-based Reasoning is that the solution to a problem provided by a CBR system can be justified showing the user the precedent case(s) used to support such a deci-sion. This form of justification is supported by DIRAS by showing for each risk assessment the cases in the discriminatory set. Moreover, DIRAS constructs a symbolic explanation with the features that are relevant in classifying a patient complication in a risk class. This symbolic explanation is close to the justification that can be provided by an expert for the same problem and may allow the user to focus on the critical features for a particular patient.

Acknowledgment

The authors thank Ana M. Monteiro for her collaboration in the acquisition of the diabetes domain knowledge. This work has been developed in the context of the SMASH project supported by the Spanish Project CICYT TIC96-1038-C04-01

References

1. Lucas PJF, Abu-Hanna A. Prognostic methods in medicine. Artificial Intelligence in Medicine

1999; 15: 105-109.

2. Kolodner JL. Case-based Reasoning. Morgan Kauffman, 1993

3. Armengol E, Plaza E. A knowledge level model of Case-based Reasoning. In Topics in Case-

based Reasoning. Lecture Notes in Artificial Intelligence, 837. S. Wess K.D Althoff and M. M.

Richter (Eds). 1994: 53-64.

4. Armengol, E. A framework for integrated learning and problem solving. Monografies de l'IIIA.

Vol 5. Institut d'Investigació en Intel.ligència Artificial Ed. Barcelona. 1997.

http://www.iiia.csic.es/Publications/monographs.html/

5. Armengol E, Plaza E. Bottom-up induction of feature terms. Machine Learning Journal. To be

published.

6. López de Mántaras R. A Distance-based Attribute Selection Measure for Decision Tree Induc-

tion. Machine Learning 1991; 6: 81-92.

7. Quinlan J.R. Induction of decision tress. Machine Learning 1986;1:81-106.

8. Verdaguer A. Patak A. Sancho J.J. Sierra C. and Sanz F. Validation of the Medical Expert Sys-

tem PNEUMON-IA. J.H. van Bemmel and A.T. McCray (Eds.) Yearbook of Medical Informat-ics 1993. IMIA's Publications, 1993: 446-461

9. Lehmann E.E, Deutsch T, Carson E.R, and S?nksen P.H. Combining rule-based resoning and

mathematical modelling in diabetes care. Artificial Intelligence in Medicine 1994; 6: 137-160.

10. Larizza C, Bellazzi R, Riva A. Temporal abstractions for diabetic patients management. Pro-

ceedings of AIME 97,E. Keravnou, C. Garbay, R. Baud and J. Wyatt eds. 1997: 319-330.

2020年公需课考试——人工智能技术及其发展趋势(93分)

单选题: 1.生物特征识别技术不包括()。(3.0分) A.体感交互 B.指纹识别 C.人脸识别 D.虹膜识别 我的答案:A√答对 2.()是人工智能的核心,是使计算机具有智能的主要方法,其应用遍及人工智能的各个领域。( 3.0分) A.深度学习 B.机器学习 C.人机交互 D.智能芯片 我的答案:B√答对 3.下列选项中,不属于生物特征识别技术的是()。(3.0分) A.步态识别 B.声纹识别 C.文本识别 D.虹膜识别 我的答案:C√答对

4.()是自然语言处理的重要应用,也可以说是最基础的应用。(3.0分) A.文本识别 B.机器翻译 C.文本分类 D.问答系统 我的答案:C√答对 5.()是指直接通过肢体动作与周边数字设备和环境进行交互。(3.0分) A.体感交互 B.指纹识别 C.人脸识别 D.虹膜识别 我的答案:A√答对 6.立体视觉是()领域的一个重要课题,它的目的在于重构场景的三维几何信息。(3.0分) A.人机交互 B.虚拟现实 C.自然语言处理 D.计算机视觉 我的答案:D√答对

7.(),中共中央政治局就人工智能发展现状和趋势举行第九次集体学习。(3.0分) A.2018年3月15日 B.2018年10月31日 C.2018年12月31日 D.2019年1月31日 我的答案:B√答对 8.()是通过建立人工神经网络,用层次化机制来表示客观世界,并解释所获取的知识,例如图像、声音和文本。(3.0分) A.深度学习 B.机器学习 C.人机交互 D.智能芯片 我的答案:A√答对 9.下列对人工智能芯片的表述,不正确的是()。(3.0分) A.一种专门用于处理人工智能应用中大量计算任务的芯片 B.能够更好地适应人工智能中大量矩阵运算 C.目前处于成熟高速发展阶段 D.相对于传统的CPU处理器,智能芯片具有很好的并行计算性能 我的答案:C√答对

炒股软件排行榜讲课教案

华尔通星级:五星 1、实时热点资讯 2、人工智能选股系统 3、股票基础知识学习视频教程(初、中、高) 4、支持线上开户,多数券商合作登录。 5、有比较多的技术指标。 6、支持模拟炒股功能 优势: 1、华尔智能选股 华尔智能选股是华尔通推出的根据对个股的量化计算、智能算法以及数据分析智能为用户选出相对优势股。分六个方面选择,包括趋势、突破、尾盘、十字星、黄金位、红三兵六种技术形态,全方位满足用户选股习惯。除此以外,华尔还能帮用户分析个股未来走势。 2、牛人圈子 华尔通设有专门的牛人圈子,资深高手汇聚,牛人会定时直播、分享经典战法、精研技术指标,更有牛人私人笔记,主要记录牛人精华内容,解读上市公司公告、季报、年报,提供最具参考的价值观点。 同花顺星级:五星 同花顺起步早,是国内第一批的企业,总体来说,优势还是有的。 优势: 1、资料齐备,你想看的信息基本上都有;

2、支持线上开户,大部分券商可以直接登录; 3、行情稳定; 4、有比较多的技术指标。 不足之处:还是偏向于传统,缺乏创新。 1.同花顺,大智慧:老牌交易软件, 功能齐全, 资讯及时, 模拟炒股为辅助, 更适合实盘操 作。 2.同花顺的最习惯最顺手,简洁、可视面积大。不过大券商基本都自己开发自家产品了。 电脑:收盘的温馨提示概况得也很实用。资讯栏信息排列得时间错乱 同花顺首页的强弱评级、赚钱效应、个股涨停分布算是挺实用的亮点。K线的指标平台各种指标都有,我觉得实用的一个是“扑捉顶底”,逃顶功能挺准的。感觉软件架构有点问题,容易假死。 雪球星级:四星 雪球最早是做美股论坛的平台,由于上线时间早,累积了大量的专业投资者。现在的雪球更像是个人投资组合的管理工具,同时也比较强调社交属性。 有趣的晒实盘的功能,增加了用户的互动。同时,雪球也推出了“买什么”产品,实质是用户自己建立的投资组合,一些投资组合是免费的,而一些是需要收费的。总的来说,在雪球上可以获得丰富的信息,而且投资者也可以非常便捷的交流。 自媒体社交,好声音打广告,股票微博类

人工智能选股之stacking集成学习

人工智能选股之stacking集成学习

本文研究导读 (4) Stacking集成学习模型简介 (5) Stacking集成学习的原理 (5) 从传统的Stacking到改进的Stacking (6) Stacking集成学习中基模型的对比和选取 (7) 相同训练数据,不同模型的对比 (7) 训练数据为72个月 (7) 训练数据为6个月 (7) 不同训练数据,相同模型的对比 (8) 模型预测值相关性分析和夏普比率分析 (9) Stacking集成学习测试流程 (10) 测试流程 (10) 模型构建 (12) Stacking模型分层回测分析 (13) 模型选股测试结果和IC值分析 (17) 对比测试1 (18) 对比测试2 (20) 对比测试3 (22) 总结和展望 (24) 附录:传统Stacking和改进Stacking的区别 (25) 传统Stacking模型的构建过程 (25) 改进Stacking模型的构建过程 (25) 风险提示 (27)

图表1: Stacking集成学习示意图 (5) 图表2:传统的Stacking集成学习 (6) 图表3:改进的Stacking集成学习 (6) 图表4:各机器学习模型相对中证500的超额收益(训练数据为72个月) (7) 图表5:各机器学习模型相对中证500的超额收益(训练数据为6个月) (8) 图表6: XGBoost各训练期长度训练所得模型相对中证500的超额收益(训练数据为6个月).. 8图表7:其他基模型预测值与XGBoost_72m预测值的相关系数 (9) 图表8:基模型夏普比率 (9) 图表9:基模型适应度指标S (9) 图表10: Stacking集成学习模型构建示意图 (10) 图表11:选股模型中涉及的全部因子及其描述 (11) 图表12: Stacking模型滚动训练过程 (12) 图表13: Stacking模型滚动测试过程 (13) 图表14:单因子分层测试法示意图 (14) 图表15: Stacking模型分层组合绩效分析(20110131~20180427) (15) 图表16: Stacking模型分层组合回测净值 (15) 图表17: Stacking模型各层组合净值除以基准组合净值示意图 (15) 图表18: Stacking模型分层组合1相对沪深300月超额收益分布图 (15) 图表19: Stacking模型多空组合月收益率及累积收益率 (15) 图表20: Stacking模型组合在不同年份的收益及排名分析(分十层) (16) 图表21:不同市值区间Stacking模型组合绩效指标对比图(分十层) (16) 图表22:不同行业Stacking模型分层组合绩效分析(分五层) (17) 图表23:对比测试1中各种模型选股指标对比(全A选股,行业中性) (18) 图表24:对比测试1中各种模型超额收益和回撤表现(全A选股,中证500行业中性,每个行业选4只个股) (19) 图表25:对比测试1中各种模型IC,IR指标 (19) 图表26:对比测试1中各种模型IC 值累积曲线 (19) 图表27:对比测试2中各种模型选股指标对比(全A选股,行业中性) (20) 图表28:对比测试2中各种模型超额收益和回撤表现(全A选股,中证500行业中性,每个行业选4只个股) (21) 图表29:对比测试2中各种模型IC,IR指标 (21) 图表30:对比测试2中各种模型IC 值累积曲线 (21) 图表31:对比测试3中各种模型选股指标对比(全A选股,行业中性) (22) 图表32:对比测试3中各种模型超额收益和回撤表现(全A选股,中证500行业中性,每个行业选4只个股) (23) 图表33:对比测试3中各种模型IC,IR指标 (23) 图表34:对比测试3中各种模型IC 值累积曲线 (23) 图表35:传统Stacking模型的构建过程 (25) 图表36:改进Stacking模型的构建过程 (26)

公需科目:2019人工智能与健康试题及答案(四)

眉山市公需科目培训:2019人工智能与健康试题及答案(四) 一、单选题 1.《“健康中国2030”规划纲要》中提到,健康是经济社会发展的()。( 2.0分) A.必然要求 B.基础条件 C.核心要义 D.根本目的 我的答案:B√答对 2.对于神经退行性疾病,()将成为主要的检测手段。(2.0分) A.X光 B.计算机断层扫描 C.核磁共振成像 D.内窥镜 我的答案:C√答对 3.在中国现有的心血管病患中,患病人数最多的是()。(2.0分) A.脑卒中 B.冠心病 C.高血压 D.肺原性心脏病 我的答案:C√答对 4.()是用电脑对文本集按照一定的标准进行自动分类标记。(2.0分) A.文本识别 B.机器翻译 C.文本分类 D.问答系统 我的答案:C√答对 5.古代把计量叫“度量衡”,其中,“度”是测量()的过程。(2.0分) A.长度 B.容积 C.温度

我的答案:A√答对 6.瓦普尼克&泽范兰杰斯理论(VC理论)的目标是很好地()事件。(2.0分) A.预测 B.统计 C.分析 D.回顾 我的答案:A√答对 7.“计算机器能够思维”的观点是由()提出来的。(2.0分) A.冯·诺依曼 B.图灵 C.诸葛亮 D.公输班 我的答案:B√答对 8.据清华原副校长施一公教授研究,中国每年有265万人死于(),占死亡人数的28%。(2.0分) A.癌症 B.心脑血管疾病 C.神经退行性疾病 D.交通事故 我的答案:A√答对 9.关于专用人工智能与通用人工智能,下列表述不当的是()。(2.0分) A.人工智能的近期进展主要集中在专用智能领域 B.专用人工智能形成了人工智能领域的单点突破,在局部智能水平的单项测试中可以超越人类智能 C.通用人工智能可处理视觉、听觉、判断、推理、学习、思考、规划、设计等各类问题 D.真正意义上完备的人工智能系统应该是一个专用的智能系统 我的答案:D√答对 10.50年前,人工智能之父们说服了每一个人:“()是智能的钥匙。”(2.0分) A.算法 B.逻辑

人工智能整合

1、人工智能诞生的标志: 1956年夏季,来自数学、心理学、神经生理学、信息论和计算机方面的十位专家,在美国达特莫斯大学召开一次历时两个月的研讨会,讨论了关于机器智能的有关问题,会上达特莫斯大学的麦卡锡提议正式采用“人工智能”一次,标志人工智能学科的正式诞生。 2、状态空间图中三元素分别代表什么? 状态空间常记为三元组:,S为初始状态的集合,F为操作的集合,G为目标状态的集合。 3、与或图的定义是? 与或图中节点代表问题:子节点为与关系的节点为与节点,子节点为或关系的节点为或节点,在与或图中无子节点的节点称为端节点。包含与或节点的图称之为与或图。 4、产生式系统推理中的三个推理定义: (1)正向推理:从事实出发,向目标方向进行推理; (2)反向推理:从目标出发,向事实方向进行推理; (3)双向推理:同时从事实和目标出发进行推理。 5、人工智能的学派: 传统划分方法:符号主义学派、连接主义学派和行为主义学派; 现代划分方法:符号智能流派、计算智能流派、群体智能流派。 6、归结策略有哪些:1、删除策略 2、支持集策略 3、线性归结策略 4、单元归结策略 5、语义归结策略祖先过滤型策略;除此之外还有锁归结策略、输入归结策略。 7、不确定性的类型:(1)随机不确定性(2)模糊不确定性(3)不完全性(4)不一致性 简答: ①人工智能的研究领域: 1、博弈 2、自动定理证明 3、专家系统 4、模式识别 5、机器学习 6、计算智能 7、自然语言处理 8、分布式人工智能 9、机器人。 ②子句集的8个步骤: (1)消去蕴含词“->”和等值词“<->”。 (2)缩小否定词的作用范围,使否定词仅作用于原子公式。 (3)变量标准化。适当改名,使得不同量词指导变量不同。 (4)消去存在量词,同时要进行变量替换。 (5)消去所有全称量词。 (6)将公式化为合取范式。 (7)适当改名,使子句之间不含同名的指导变量。 (8)消去合取词,以子句为元素组成一个集合S。 1、状态空间图:状态、操作、状态空间图、求解 2、状态空间图的盲目搜索算法的概念和步骤:深度优先、广度优先(教材30-32页) 3、状态空间图的启发式搜索算法的概念:以启发性知识为导航的搜索就是启发式搜索。 按照考察节点的选择范围不同,算法分为全局择优和局部择优两种。 4、A算法:启发式搜索算法中同时考虑初始节点到当前节点已经付出的代价和当前节点到目标节点的代价,即引入估价函数f(x)=g(x)+h(x) 5、 A*算法:A*算法是一种启发式搜索方法,搜索时对扩展节点的选择方法做了一些限制。要求根据估价函数 f(x)=g(x)+h(x) 对OPEN表中的节点进行排序,并且要求启发函数 h(x) 是 h*(x) 的一个下界,即 h(x)<=h*(x)。h*(x) 是从x节点到目标节点的最小代价路径上的代价。 A* 算法和A算法的区别就是A算法不要求启发函数h(x) 是 h*(x) 的一个下界,即不限制条件h(x)<=h*(x)。A*算法具有可采纳性(如果问题有解,该算法一定能够在有限步内找到一条最优解)、单调性(启发函数值单调递增)、信息性(启发函数的值越大,搜索效率越高) 6、与或图:与或图中节点代表问题:子节点为与关系的节点为与节点,子节点为或关系的节点为或节点,在与或图中无子节点的节点称为端节点。包含与或节点的图称之为与或图。

示意图画法

怎样画示意图题 一、显示空间关系 1.火车长100m ,车头距离桥头200m ,桥长200m ,火车从静止开始以a =1m/s 2的加速度运动,求火车过桥经历的时间。 2.长5.0m 的铁链悬于O 点,O 点下方距离铁链下方15m 处有一个(偏离O 点正下方少许)钉子。求铁链无初速释放后经过钉子的时间是多长?(g 取10m/s ) 3.1999年高考题 在光滑水平面上有一质量m =1.0×10-3 Kg 、电量q =1.0×10-10 C 的带正电小球,静止在O 点。以O 点为原点,在该水平面内建立直角坐标系Oxy 。现突然加一沿x 轴正方向、场强大小E =2.0×106V/m 的匀强电场,使小球开始运动。经过1.0s ,所加电场突然变为沿y 轴正方向,场强大小仍为E =2.0×106 V/s 匀强电场。再经过1.0s ,所加电场又突然变为另一个匀强电场,使小球在此电场作用下经1.0s 速度变为零。求此电场的方向及速度变为零时小球的位置。 4.2006年理综Ⅰ卷第23题 天空有近似等高的浓云层。为了测量云层的高度,在水平地面上与观测者的距离为d =3.0km 处进行一次爆炸,观测者听到由空气直接传来的爆炸声和由云层反射来的爆炸声时间上相差Δt =6.0s 。试估算云层下表面的高度。已知空气中的声速v =1 3 km/s 。 5.2007年理综Ⅰ卷第23题 甲乙两运动员在训练交接棒的过程中发现:甲经短距离加速后能保持9m /s 的速度跑完全程:乙从起跑后到接棒前的运动是匀加速的。为了确定乙起跑的时机,需在接力区前适当的位置设置标记。在某次练习中,甲在接力区前013.5m s =处作了标记,并以9m /s v =的速度跑到此标记时向乙发出起跑口令。乙在接力区的前端听到口令时起跑,并恰好在速度达到与甲相同时被甲追上,完成交接棒,已知接力区的长度为L =20m. 求:(1)此次练习中乙在接棒前的加速度a 。 (2)在完成交接棒时乙离接力区末端的距离. 二、把立体关系转化为平面关系: 6.如图所示,abcd 是一竖直的矩形导线框,线框面积为S ,放在磁感应强度为B 的均匀水平磁场中.ab 边在水平面内且与磁场方向成60?角.则通过导线框的磁通量等于 ( ) (A)BS (B) 12BS (C) 2 2BS (D) 32 BS

四句话就让你彻底明白人工智能都是什么!

四句话就让你彻底明白人工智能都是什么! 第一句:AI顾名思义就是英文单词Artificial intelligenc,即人工智能。 其实人工智能并不是什么触不可及的东西,包括苹果Siri、百度度秘、微软小冰等智能助理和智能聊天类应用以及美图秀秀的自动美化功能,都属于人工智能。甚至一些简单的,套路固定的资讯类新闻,也是由人工智能来完成的。 当然,现在的主流搜索引擎以及翻译技术也都在尝试,利用人工智能来为广大网友提供更为精准的搜索服务。至于以实物存在的人工智能,当属现在物流仓库的小黄机器人了。他们正代替人类完成繁重的商品摆放、整理,快速出库、入库等操作。 第二句:现在人工智能并没有发展到像电影中的机器人一样,那么高智能化的程度,今天的家庭机器人还远无法像大家奢望的那样,以人形外貌出现在主人面前。现在的人工智能分三个级别:弱人工智能、强人工智能、超人工智能。 1、弱人工智能 也称限制领域人工智能或应用型人工智能,指的是专注于且只能解决特定领域问题的人工智能。毫无疑问,今天我们看到的所有人工智能算法和应用都属于弱人工智能。Alpha Go 其实也是一个弱人工智能。 2、强人工智能 强人工智能又称通用人工智能或完全人工智能,指的是可以胜任人类所有工作的人工智能。一个可以称得上强人工智能的程序,大概需要具备以下几方面的能力:存在不确定因素时进行推理,使用策略,解决问题,制定决策的能力;知识表示的能力,包括常识性知识的表示能力;规划能力;学习能力;使用自然语言进行交流沟通的能力;将上述能力整合起来实现既定目标的能力。 3、超人工智能 假设计算机程序通过不断发展,可以比世界上最聪明、最有天赋的人类还聪明,那么由此

关于人工智能技术及其发展趋势2020考试答案

人工智能技术及其发展趋势 1.关于专用人工智能与通用人工智能,下列表述不当的是()。(10.0分) A.人工智能的近期进展主要集中在专用智能领域 B.专用人工智能形成了人工智能领域的单点突破,在局部智能水平的单项测试中可以超越人类智能 C.通用人工智能可处理视觉、听觉、判断、推理、学习、思考、规划、设计等各类问题 D.真正意义上完备的人工智能系统应该是一个专用的智能系统 我的答案:D√答对 2.()是通过建立人工神经网络,用层次化机制来表示客观世界,并解释所获取的知识,例如图像、声音和文本。(10.0分) A.深度学习 B.机器学习 C.人机交互 D.智能芯片 我的答案:A√答对 3.下列选项中,不属于生物特征识别技术的是()。(10.0分) A.步态识别 B.声纹识别 C.文本识别

D.虹膜识别 我的答案:C√答对 4.()是自然语言处理的重要应用,也可以说是最基础的应用。(10.0分) A.文本识别 B.机器翻译 C.文本分类 D.问答系统 我的答案:C√答对 1.一般说来,人工智能技术包括()。(10.0分)) A.深度学习、机器学习 B.计算机视觉、自然语言处理 C.人机交互、生物信息技术、智能芯片 D.虚拟现实/增强现实、机器人技术 我的答案:ABCD√答对 2.指纹识别是通过()等物理传感器获取指纹图像,经过数据处理进行分析判别。(10.0分)) A.光 B.电 C.力 D.热

我的答案:ABCD √答对 1.机器学习是一门多领域交叉学科,涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。(10.0分) 我的答案: 正确 √答对 2.卷积神经网络是一种常用来处理具有网格结构拓扑数据的神经网络,如处理时序数据和图像数据等,广泛应用于人脸识别、物品识别等领域。(10.0分) 我的答案: 正确 √答对 3.人工智能是科学交叉的结果,它由不同领域多学科综合发展而来。(10.0分) 我的答案: 正确 √答对 4.机器人一般由执行机构、驱动装置、检测装置、控制系统和复杂机械等组成,涉及到控制论、机械电子、计算机、材料、仿生等学科,在工业、医学、农业、建筑业甚至军事等领域中均有重要用途。(10.0分) 我的答案: 正确 √答对 1、最困难的事就是认识自己。20.7.57.5.202008:2708:27:50Jul-2008:27 2、自知之明是最难得的知识。二〇二〇年七月五日2020年7月5日星期日 3、越是无能的人,越喜欢挑剔别人。08:277.5.202008:277.5.202008:2708:27:507.5.202008:277.5.2020 4、与肝胆人共事,无字句处读书。7.5.20207.5.202008:2708:2708:27:5008:27:50 5、三军可夺帅也。Sunday, July 5, 2020July 20Sunday, July 5, 20207/5/2020 6、最大的骄傲于最大的自卑都表示心灵的最软弱无力。8时27分8时27分5-Jul-207.5.2020 7、人生就是学校。20.7.520.7.520.7.5。2020年7月5日星期日二〇二〇年七月五日 亲爱的用户: 烟雨江南,画屏如展。在那桃花盛开的地方,在这醉 人芬芳的季节,愿你生活像春天一样阳光,心情像桃花一 样美丽,感谢你的阅读。

关于机械电子工程与人工智能整合思路构建的研究

关于机械电子工程与人工智能整合思路构建的研究 摘要:机械、电子以及网络信息等新兴科技的发展,促进了我国机械工程技术的进步,在机械电子工程阶段升级的基础上,将进一步融合先进的人工智能技术,从而提高机械电子工程的数字化和自动化水平,并向智能化方向快速发展。在机械电子工程中融合人工智能技术,对于推动人类社会的进步具有十分重要的意义,同时也是机械工程的一次重要变革。因此相关研究人员应加大对整合人工智能和机械电子工程思路的研究,促进二者深入融合与发展。 关键词:机械电子工程;人工智能;整合思路 随着各种高新技术的快速发展,电子技术在传统机械工程中广泛应用,实现机电一体化的机械电子工程,提高机械工程的现代化水平。同时,以网络信息技术为基础的人工智能技术目前已经成为科技发展的主要趋势之一,因此在机械电子工程中也要加强与人工智能技术的融合,提高机械电子工程的自动化和数字化水平,促使机械电子工程加快智能化发展的速度,从而更好地适应时代发展的需要,为我国经济建设和社会生活提供更加便捷高效的智能化机械电子产品,也能够为机械电子企业创造更大的经济效益。机械电子工程与人工智能的整合将促使社会生产力发生重大的变革,相关企业和研究人员应加大机械电子工程与人工智能整合思路构建的研究力度。 1 概述人工智能以及机械电子工程 1.1 概述机械电子工程特点 随着科技的发展,目前机械电子工程已经成为融合机械工程技术和电子工程技术的综合性系统工程,计算机技术是机械电子工程中的重要组成部分,而电子工程技术和计算机技术与传统机械工程技术的结合极大丰富了机械电子产品的功能,同时也使结构更加简洁,扩大机械电子产品的应用范围,因此机械电子工程具有广阔的发展前景[1]。 1.2 概述人工智能技术 随着计算机以及网络信息技术的发展,目前人工智能技术已经进入到快速发展阶段。人工智能技术在数据处理分析等方面功能更加强大,在机械制造等工业生产中可以完成模型构建,并進行故障预警以及故障排除等,对于全面提高机械电子工程的现代化水平和智能化程度具有十分重要的作用。 2 关于整合机械电子工程以及人工智能的思路构建研究 2.1 分析人工智能技术与机械电子工程之间的关系 传统的机械电子工程在稳定性方面有所欠缺,特别是主要采用的数学方程推

装配示意图的画法装配示意图常见画法

装配示意图的画法 装配示意图是用线条和符号来表示零件间的装配关系和装配体工作方式的一种工程简图,它主要表明部件中各零件的相对位置、装配连接关系和运转情况,以确保画装配图和重新装配工作的顺利进行。装配示意图也是绘制装配图时的重要参考资料。 (1)装配示意图的常用符号。 装配示意图用线条和符号来表示零件间的装配关系,但目前装配示意图的符号还没有统一的规定。在工程实践中,人们创造了一些常用零件的符号,其中一些符号被广泛采用,已有约定俗成的趋势。常用的符号见如下附表,供测绘时参考。 (2)装配示意图的两种常见画法 装配示意图的画法也没有统一的规定。通常,图上各零件的结构形状和装配关系,可用较少的线条形象地表示,简单的甚至可以只用单线条来表示。目前,,较为常见的有“单线+符号”和“轮廓+符号”两种画法。 A,用“单线+符号”画法画装配示意图, “单线+符号”画法是将结构件用线条来表示,对装配体中的标准件和常用件用符号来表示的一种装配示意图画法。用这种画法绘制装配示意图时,两零件间的接触面应按非接触面的画法来绘制。 图l—2所示为球阀的轴测图、装配图及装配示意图。图l—2(c)中零件9和零件14,零件10、11和12之间都是接触表面,在图中要用两条线来表示。其中所有的非标准件都是用单线来表示的。 B,用“轮廓+符号”画法画装配示意图 装配示意图的另一种画法是“轮廓+符号”画法。这种画法是画出部件中一些较大零件的轮廓,其他较小的 零件用单线或符号来表示。 图l—3所示为螺旋千斤顶的轴测图、装配图和装配示意图。在图l—3(c)中,千斤顶外壳、顶盖的画法采 用了轮廓画法。

虎钳装配示意图

通用人工智能与机器的意识问题

通用人工智能与机器的意 识问题 南开大学黄彧

?意识是什么——意识的哲学思考??Conscious AGI Agent是否可能? ?Conscious AGI Agent如何构建自己的身体??Conscious AGI Agent是怎样与环境互动的? ?如何理解AGI Agent中的时间问题??Conscious AGI Agent实现自由意志的途径是什么??AGI Agent如何具有主观经验? ?AGI Agent中意识和创造力有何种联系?

?Antonio Chella University of Palermo 研究领域:Artificial Intelligence; Oncology; Machine Learning & Pattern Recognition; Cardiology; Multimedia antonio.chella@unipa.it ?Riccardo Manzotti Iulm University of Milan 研究领域:Artificial Intelligence; Computer Vision; Human-Computer InteractionNetworks & Communications; Neuroscience riccardo.manzotti@iulm.it One's conscious experience of an object is identical with the object one experiences

对在设计有意识的智能体时面临的主要问题,进行回顾。

2019年人工智能系列研究:机器学习中的随机数

2019年人工智能系列研究:机器学习中的随机数

正文目录 本文研究导读 (4) 机器学习中的随机数 (5) 从计算机中的随机数生成谈起 (5) 数据集的随机划分 (5) 优化算法中的随机数 (6) 赋予参数随机初始值 (6) 随机梯度下降 (6) 集成学习中的随机数 (8) 神经网络中的随机数 (9) Python环境下如何设置随机数种子 (10) 机器学习选股模型随机性的来源 (11) 方法 (12) 人工智能选股模型测试流程 (12) 全连接神经网络模型参数设定 (14) 单因子测试 (14) 回归法和IC值分析法 (14) 分层回测法 (15) 结果 (16) 模型性能 (16) 回归法和IC值分析法 (17) 分层测试法 (18) 不同随机性来源的横向比较 (20) 总结 (22) 风险提示 (23) 图表目录 图表1:机器学习中随机数所涉及的环节、作用和代表模型 (5) 图表2:二元损失函数示意图 (7) 图表3:损失函数为凸函数(左)和非凸函数(右) (7) 图表4:梯度下降法(左)和随机梯度下降法(右) (7) 图表5:Bootstrap重采样示意图 (8) 图表6:Bagging并行集成方法示意图 (9) 图表7:Dropout方法示意图 (10) 图表8:Python常用机器学习包中随机数种子参数设置方法 (11) 图表9:keras包(tensorflow作为后端)设置随机数种子代码实例 (11) 图表10:机器学习选股模型随机性的可能来源和对应的考察方式 (11)

图表11:人工智能选股模型测试流程示意图 (12) 图表12:年度滚动训练示意图 (12) 图表13:选股模型中涉及的全部因子及其描述 (13) 图表14:模型历年滚动训练最优超参数 (14) 图表15:2011~2018年四种模型样本外平均正确率分布 (16) 图表16:2011~2018年四种模型样本外平均AUC分布 (16) 图表17:2018年四种模型样本外平均正确率分布 (16) 图表18:2018年四种模型样本外平均AUC分布 (16) 图表19:2011~2018年四种模型平均t 值分布 (17) 图表20:2011~2018年四种模型平均|t|值分布 (17) 图表21:2011~2018年四种模型平均因子收益率分布 (17) 图表22:2011~2018年四种模型平均Rank IC分布 (17) 图表23:2011~2018年逻辑回归模型累积Rank IC及波动情况 (18) 图表24:2011~2018年XGBoost模型累积Rank IC及波动情况 (18) 图表25:2011~2018年随机森林模型累积Rank IC及波动情况 (18) 图表26:2011~2018年全连接神经网络模型累积Rank IC及波动情况 (18) 图表27:2011~2018年四种模型多空组合年化收益率分布 (18) 图表28:2011~2018年四种模型多空组合夏普比率分布 (18) 图表29:2011~2018年四种模型Top组合年化收益率分布 (19) 图表30:2011~2018年四种模型Top组合夏普比率分布 (19) 图表31:2011~2018年逻辑回归模型多空组合净值及波动情况 (19) 图表32:2011~2018年XGBoost模型多空组合净值及波动情况 (19) 图表33:2011~2018年随机森林模型多空组合净值及波动情况 (19) 图表34:2011~2018年全连接神经网络模型多空组合净值及波动情况 (19) 图表35:2011~2018年逻辑回归模型多空组合平均和最优最差净值 (20) 图表36:2011~2018年XGBoost模型多空组合平均和最优最差净值 (20) 图表37:2011~2018年随机森林模型多空组合平均和最优最差净值 (20) 图表38:2011~2018年全连接神经网络多空组合平均和最优最差净值 (20) 图表39:XGBoost模型四种随机性来源比较 (21)

2019年度人工智能与健康考试答案

2019年度人工智能与健康考试答案91分 1.《“健康中国2030”规划纲要》中提到,健康是经济社会发展的()。( 2.0分) A.必然要求 B.基础条件 C.核心要义 D.根本目的 我的答案:B√答对 2.对于神经退行性疾病,()将成为主要的检测手段。(2.0分) A.X光 B.计算机断层扫描 C.核磁共振成像 D.内窥镜 我的答案:C√答对 3.在中国现有的心血管病患中,患病人数最多的是()。(2.0分) A.脑卒中 B.冠心病 C.高血压

D.肺原性心脏病 我的答案:C√答对 4.()是用电脑对文本集按照一定的标准进行自动分类标记。(2.0分) A.文本识别 B.机器翻译 C.文本分类 D.问答系统 我的答案:C√答对 5.古代把计量叫“度量衡”,其中,“度”是测量()的过程。(2.0分) A.长度 B.容积 C.温度 D.轻重 我的答案:A√答对 6.瓦普尼克&泽范兰杰斯理论(VC理论)的目标是很好地()事件。(2.0分) A.预测 B.统计 C.分析 D.回顾 我的答案:A√答对

7.“计算机器能够思维”的观点是由()提出来的。(2.0分) A.冯·诺依曼 B.图灵 C.诸葛亮 D.公输班 我的答案:B√答对 8.据清华原副校长施一公教授研究,中国每年有265万人死于(),占死亡人数的28%。(2.0分) A.癌症 B.心脑血管疾病 C.神经退行性疾病 D.交通事故 我的答案:A√答对 9.关于专用人工智能与通用人工智能,下列表述不当的是()。(2.0分) A.人工智能的近期进展主要集中在专用智能领域 B.专用人工智能形成了人工智能领域的单点突破,在局部智能水平的单项测试中可以超越人类智能 C.通用人工智能可处理视觉、听觉、判断、推理、学习、思考、规划、设计等各类问题 D.真正意义上完备的人工智能系统应该是一个专用的智能系统 我的答案:D√答对

人工智能伦理学慕课题库

人工智能伦理学慕课题库 1.1人工智能的历史 1. [多选题] 对人工智能常见的误解有哪些?( ) A.人工智能就是机器学习 B.机器学习只是人工智能中的一个方向 C.人工智能最近十年受到深度学习的驱动较多 D.人工智能就是深度学习 我的答案:AD 2. [判断题] 哲学思维对于人工智能的重要性表现在,哲学所强调的批判性思维有助于认清人工智能发展中的问题。( ) 我的答案:对 3. [判断题] 深度学习在人工智能领域的表现并不突出。( ) 我的答案:错 1.2符号人工智能 1. [单选题] 人工智能作为一门学科的建立时间是( )。 A.1956年 B.1930年 C.1960年 D.1952年 我的答案:A 2. [单选题] 人工智能与计算机学科的关系是( )。 A.计算机学科的主要驱动力是人工智能研究 B.计算机是人工智能研究的一个领域 C.人工智能是计算机学科的一个分支 D.人工智能与计算机学科没有联系 我的答案:C 3. [单选题] 计算机之父是( )。 A.约翰·麦卡锡 B.艾伦·图灵 C.赫尔伯·西蒙 D.马文·明斯基 我的答案:B 4. [判断题] 符号AI是将人的思维通过逻辑语言制成流形图让计算机去执行。( ) 我的答案:对 5. [判断题] 通用问题求解器需要寻找全局最优解。( ) 我的答案:错 1.3人工神经网络

1. [单选题] ( )是现在新出现的人工智能的研究方向。 A.深度学习 B.人工神经元网络 C.贝叶斯网络 D.类脑人工智能 我的答案:D 2. [单选题] 深度学习中的“深度”是指( )。 A.计算机理解的深度 B.中间神经元网络的层次很多 C.计算机的求解更加精准 D.计算机对问题的处理更加灵活 我的答案:B 3. [多选题] 人工神经元网络与深度学习的关系是( )。 A.人工神经元网络是深度学习的前身 B.深度学习是人工神经元网络的一个分支 C.深度学习是人工神经元网络的一个发展 D.深度学习与人工神经元网络无关 我的答案:AC 4. [判断题] 符号AI不是人工智能的正统。( ) 我的答案:错 5. [判断题] 相比于人工神经元网络和深度学习,类脑人工智能对人类大脑的神经回路具有更深入的了解。( ) 我的答案:对 1.4框架问题 1. [单选题] 深度学习的实质是( )。 A.推理机制 B.映射机制 C.识别机制 D.模拟机制 我的答案:B 2. [判断题] 计算机具有触类旁通的能力,可以根据具体语境对事件进行分类。( ) 我的答案:错 3. [判断题] 人工神经元网络会遭遇“框架问题”。( ) 我的答案:错 4. [判断题] 推理的本质是在信息不足的情况下能够最大程度的得到最靠谱 的结论。( ) 我的答案:对1. [单选题] 深度学习的实质是( )。 A.推理机制 B.映射机制 C.识别机制 D.模拟机制 我的答案:B

量化系统公开课

散户不追高,韭菜变镰刀低吸富三代,追高悔一生 股票应该怎么看?先看大趋势、再看小趋势、看对大方向、做好小波段、历史不会重演、但会惊人的相似! 量化 实战派只讲实战,潜伏涨停战法:人工智能交易选股、长短结合、专一、专注、专业 全网公开智能选股检验技术、主动买套被动盈利、潜伏涨停板战法、买在黄金套、首家公开透明、真实可靠实战教学 量化教学 明确直观易学易懂 注重实战眼见为实 重执行守纪律 智能选股仓位管理 成功的思考:希望量化团队到牛市吃一波10倍资金增长1、学会智能交易选股策略,一杯咖啡时间选股;2、学会仓位资金管理交易方法;3、学会简单基本面了解方法与技巧; 4、学会如何调整心态坚决不追高; 5、学会股价什么位置叫高与低,远离追高。 《量化交易密决》 好的交易系统盈利能力看长期稳定性,不是短期 下楼梯、走平台、买转角、起主升一次学习、受益终生 量化老师粉丝QQ 群 514483762 领干货课件试听 好票锁仓反复做,买在黄金套,潜伏涨停板战法!一、为机构提供策略编写、二次软件开发,只讲实战。一、每天公开股票实战,股票实战技术经验免费分享;三、每天公开期货实战,全自动化期货交易系统分享;风险提醒:大家千万不要盲目跟进,不懂原理的很容易造成亏损,因为交易有明确的交易计划,有意向学习的可以加盟 量化太极院

散户不追高,韭菜变镰刀炒股大智慧,格局赚大钱 一学就会、成就高手、人工智能选股、不是纸上谈兵、而是有效大数据分析 量化系统:专抄机构底、让交易变的简单、让小白成为高手!一直被模仿一、短线做法如何提高成功率高? 怎样明确定义看长做短? 二、晚上8点30分时间形态学技术与方法三、勾选上课通知我,我上课时会通知大家 拜师嘉年华,拜师有优惠 促销活动,最大力度,到12月份截止,月徒送多1个月,如果学员达到5位,多送2个月共3个月,欢迎咨询 !加 量化老师粉丝QQ 群 514483762 领学员课件试听 我们在低位低估的时候买入未来上涨的概率会更大,盈利赚到的钱也会越多,相反我们在高位高估的时候投资,未来下跌的概率会更大,赚到的钱也会大大减少,甚至亏损。实际投资过程中,很多人无法 理解这一点。我们要尽量做胜率高、赔率低,对我们有利的决策,也就是低位低估的时候投资。 关注量化太极院、低吸富三代、追高悔一生、佛渡有缘人、简单通透、易学易懂,真实有效、稳定套利、太极阴阳理论 量化教学 明确直观易学易懂 注重实战眼见为实 重执行守纪律 智能选股仓位管理

2019年人工智能技术分析报告

2019年人工智能技术分析报告

正文目录 本文研究导读 (4) 时序交叉验证的改进 (5) K 折和时序交叉验证 (5) 改进思路1——更合理的基线模型 (6) 改进思路2——更精细的切分方法 (7) 方法 (8) 人工智能选股模型测试流程 (8) 单因子测试 (10) 回归法和IC 值分析法 (10) 分层回测法 (10) 结果 (11) 最优超参数 (11) 模型性能 (12) 单因子测试 (13) 构建策略组合及回测分析 (15) 总结和讨论 (18) 附录:分组时序交叉验证的代码实现 (19) 修改model_selection 包的_split.py (19) 修改model_selection 包的__init__.py (20) 主函数中调用GroupTimeSeriesSplit 类 (21) 图表目录 图表1:K 折交叉验证示意图(K=5) (5) 图表2:时序交叉验证示意图(折数=5) (5) 图表3:新基线模型1:训练集折半的K 折交叉验证示意图(K=5) (6) 图表4:新基线模型2:乱序递进式交叉验证示意图(折数=5) (6) 图表5:分组时序交叉验证示意图(折数=5) (7) 图表6:新基线模型3:乱序分组递进式交叉验证示意图(折数=5) (7) 图表7:本文测试的六种交叉验证方法汇总 (7) 图表8:人工智能选股模型测试流程示意图 (8) 图表9:年度滚动训练示意图 (8) 图表10:选股模型中涉及的全部因子及其描述 (9) 图表11:选股模型超参数和调参范围 (10) 图表12:模型历年滚动训练最优超参数 (11) 图表13:六种交叉验证方法模型性能对比(回测期20110131~20190131) (12)

2020人工智能试题与答案

1.()是普遍推广机器学习的第一人。( 2.0分) A.约翰·冯·诺依曼 B.约翰·麦卡锡 C.唐纳德·赫布 D.亚瑟·塞缪尔 我的答案:C√答对 2.当我们需要寻求健康咨询服务时,应该拨打的热线电话是()。(2.0分) A.12315 B.12301 C.12345 D.12320 我的答案:D√答对 3.()由于产品全球化市场竞争加剧和信息技术革命的推动,围绕提高制造业水平的新概念和新技术不断涌现,在此背景下,将新兴的人工智能技术应用于制造领域使“智能制造”的概念孕育而生,并促进了智能制造技术和智能制造系统的研究。(2.0分) A.20世纪70年代 B.20世纪80年代 C.20世纪90年代 D.21世纪初 我的答案:C√答对 4.我国于()年发布了《国务院关于印发新一代人工智能发展规划的通知》。(2.0分) A.2016 B.2017 C.2018 D.2019 我的答案:B√答对

5.在农业领域的()环节,智能的农业机器人可以利用图像识别技术获取农作物的生长状况,判断哪些杂草需要清除,判断哪里需要灌溉、施肥、打药,并立即执行。(2.0分) A.产前 B.产中 C.产后 D.全程 我的答案:B√答对 6.()是人工智能发展的硬道理,没有它的人工智能是没有用的。(2.0分) A.数据 B.应用 C.逻辑 D.算法 我的答案:B√答对 7.新生儿的正常脉搏为每分钟()次。(2.0分) A.60~80 B.70~90 C.80~100 D.90~120 我的答案:D√答对 8.2016年8月,日本电视台报道称,东京大学医学研究所通过运用IBM的人工智能平台Watson仅 用10分钟就诊断出了资深医师难以判别出来的()。(2.0分) A.甲状腺癌 B.胰腺癌 C.淋巴癌 D.白血病 我的答案:C×答错

从专用人工智能迈向通用人工智能

论坛Forum 本次大会主题是“智联世界,无限可能”,我们感觉到AI领域热度不减,突破连连。腾讯建立的四大AI 实验室涵盖了从全面的基础研究到多种应用开发,涵盖机器人、量子计算、边缘计算、IoT物联网等等,打造面向未来的科技引擎。 第一,通用人工智能发展趋势越来越清晰。AI向人类社会继续靠近,从专才向通才发展,实现从专用人工智能向通用人工智能的跨越发展,这也是下一阶段的重要趋势。其中重要的一步就是“人工+智能”向自主智能转化,以AlphaGo的后续版本AlphaZero为代表,人的干预减少,机器自我学习能力大大增强。从2016年起,腾讯的AI Lab连续多年打造世界冠军围棋AI,随后和王者荣耀合作开发了AI绝悟,绝悟不依赖人 类的经验,它从零到一自学成才, 目前绝悟已经达到了职业水平,这 是从专用人工智能迈向通用人工智 能的重要一步。 虚拟世界作为真实世界的模拟 和仿真,一直是检验和提升AI能 力的试金石,而复杂的虚拟环境被 业界认为是攻克AI的难题,也就 是关键一步。如果我们在模拟真实 世界的虚拟游戏中,AI学会跟人 一样进行快速分析决策和行动,那 么AI就能够执行更加困难和更加 复杂的任务。举一个例子,大家知 道可观测宇宙原子的数量是10的 80次方,到了王者荣耀,决策的复 杂度可以高达10的2万次方,在如此 复杂的虚拟环境,如果AI都能表现 得类似或者超越人类,在真实世界里 要实现通用就非常值得期待了。这个 得益于我们游戏的物理引擎具有强大 的仿真能力,让AI有了逼真的训练 效果。目前,我们在人工智能研究的 人工模拟方面表现也是全球比较领先 的。我们从绝艺到绝悟,不断累积强 大的算法,让AI有了超强大脑。 第二,在AI应用方面,AI与各 行各业日益融合,人工智能发展将会 给我们带以“AI+”为标志的普惠型 智能社会。人工智能产业未来十年在 我国将进入高速发展期,目前AI和 各个行业结合之后形成的智慧制造、 智慧安防、智慧零售、智慧医疗、智 慧交通等方面全方位的行业解决方案 正在全国落地生根。 智慧医疗是我们 最关注的重点,我们建设了医疗营销 人工智能平台,辅助医生诊断多种癌 症和疾病,可以对700多种疾病进行 识别和预测,目前已经成为了国家级 的标杆项目。我们还通过互联网医院、 电子健康卡等专业医学科普项目构成 了医疗领域AI应用的小生态。 第三,AI治理方面,以科技向 善引领全方位治理,确保AI可知、 可用、可控、可靠,这是我在去年大 会当中提出来的。过去一年我们看到 AI治理的紧迫性越来越高,我们也 在思考。我们今年提出了科技向善, 作为腾讯的新的使命,探索AI与人 的探索之道。 今年6月,国家提出“发展负责 任的人工智能”,这是我国首次发布 发展人工智能治理的原则。智能科技 要维护伦理道德底线,实现行业和企 业的伦理自律准则。政府、企业、社 会在AI治理大原则上有高度共识。 另外,我还想强调加强全球治理 和合作是AI发展当中必不可少的一 环,今天没有哪一个国家拥有全部的 资源、技术和能力。产业割裂、技术 脱钩会损害人类的长期利益,面对种 种矛盾,我们应该拿出智慧和胸怀, 努力跨越这些壁垒。 最后,感谢大会组织方以卓越 的远见和热情,把政府、科技、人 文、教育、企业、投资等各个方面的 力量凝聚在一起探讨AI的发展前景。 (本文系作者在2019世界人工 智能大会上的演讲) 从专用人工智能迈向通用人工智能 ◎ 腾讯董事会主席兼首席执行官 马化腾 9 2019.09

相关文档