MindMap Gallery Artificial Intelligence Basics
The basics of artificial intelligence are some relatively traditional methods that do not involve deep learning. For deep learning, you can see another mind map of mine. This article summarizes an overview of artificial intelligence, deterministic knowledge systems, uncertainty reasoning, intelligent search technology, etc.
Edited at 2024-02-04 00:47:36Mappa mentale per il piano di inserimento dei nuovi dipendenti nella prima settimana. Strutturata per giorni: Giorno 1 – benvenuto, configurazione strumenti, presentazione team. Secondo giorno – formazione su policy aziendali e obiettivi del ruolo. Terzo giorno – affiancamento e primi task guidati. Il quarto giorno – riunioni con dipartimenti chiave e feedback intermedio. Il quinto giorno – revisione settimanale, definizione obiettivi a breve termine e integrazione culturale.
Mappa mentale per l’analisi della formazione francese ai Mondiali 2026. Punti chiave: attacco stellare guidato da Mbappé, con triplice minaccia (profondità, taglio, sponda). Criticità: centrocampo poco creativo – la costruzione offensiva dipende dagli attaccanti che arretrano. Difesa solida (Upamecano, Saliba, Koundé). Portiere Maignan. Variabili: gestione infortuni e condizione fisica dei big. Ideale per scout, giornalisti e tifosi.
Mappa mentale per l’analisi della formazione francese ai Mondiali 2026. Punti chiave: attacco stellare guidato da Mbappé, con triplice minaccia (profondità, taglio, sponda). Criticità: centrocampo poco creativo – la costruzione offensiva dipende dagli attaccanti che arretrano. Difesa solida (Upamecano, Saliba, Koundé). Portiere Maignan. Variabili: gestione infortuni e condizione fisica dei big. Ideale per scout, giornalisti e tifosi.
Mappa mentale per il piano di inserimento dei nuovi dipendenti nella prima settimana. Strutturata per giorni: Giorno 1 – benvenuto, configurazione strumenti, presentazione team. Secondo giorno – formazione su policy aziendali e obiettivi del ruolo. Terzo giorno – affiancamento e primi task guidati. Il quarto giorno – riunioni con dipartimenti chiave e feedback intermedio. Il quinto giorno – revisione settimanale, definizione obiettivi a breve termine e integrazione culturale.
Mappa mentale per l’analisi della formazione francese ai Mondiali 2026. Punti chiave: attacco stellare guidato da Mbappé, con triplice minaccia (profondità, taglio, sponda). Criticità: centrocampo poco creativo – la costruzione offensiva dipende dagli attaccanti che arretrano. Difesa solida (Upamecano, Saliba, Koundé). Portiere Maignan. Variabili: gestione infortuni e condizione fisica dei big. Ideale per scout, giornalisti e tifosi.
Mappa mentale per l’analisi della formazione francese ai Mondiali 2026. Punti chiave: attacco stellare guidato da Mbappé, con triplice minaccia (profondità, taglio, sponda). Criticità: centrocampo poco creativo – la costruzione offensiva dipende dagli attaccanti che arretrano. Difesa solida (Upamecano, Saliba, Koundé). Portiere Maignan. Variabili: gestione infortuni e condizione fisica dei big. Ideale per scout, giornalisti e tifosi.
Artificial Intelligence Basics
Artificial Intelligence Overview
AI basic concepts
smart concept
natural intelligence
definition
Refers to the power and behavioral abilities of humans and some animals
natural human intelligence
It is the comprehensive ability of human beings in understanding the objective world that is manifested by thinking processes and mental activities.
Different views and hierarchies of intelligence
View
theory of mind
Intelligence comes from thinking activities
knowledge threshold theory
Intelligence depends on applicable knowledge
evolutionary theory
Intelligence can be achieved by gradual evolution
Hierarchy
Characteristic capabilities included in intelligence
Perception
memory and thinking skills
learning and adaptability
capacity
Artificial intelligence concept
explain
Use artificial methods to achieve intelligence on machines
Study how to construct intelligent machines or systems, and simulate and extend artificial intelligence
Turing test
Artificial Intelligence Research Goals
The emergence and development of AI
gestation period
formative period
knowledge application period
From school separation to synthesis
Machine learning and deep learning lead the way
Basic content of AI research
The subject position of artificial intelligence
The intersection of natural sciences and social sciences
Core: Thinking and Intelligence
Basic subjects: mathematics, thinking science, computer
Interdisciplinary research with brain science and cognitive science
Research on methods and technologies of intelligent simulation
machine perception
Vision
hearing
machine thinking
machine learning
machine behavior
Different schools of AI research
symbolism
mathematical logic
knowledge engineering
connectionism
bionics
Artificial neural networks
Behaviorism
cybernetics
deterministic knowledge system
An overview of deterministic knowledge systems
knowledge representation
knowledge definition
Types of knowledge
Require
expressive ability
Correct and effective representation
availability
Conducive to effective reasoning
Organizability and maintainability
can be organized into knowledge structures in some way
understandability and achievability
Easy to read, easy to understand, easy to obtain and easy to implement
display method
type
declarative knowledge
Knowledge itself and the process of use are separated from each other
transitional knowledge
The knowledge itself and the process of use are combined
basic method
unstructured approach
predicate logic
production
structured approach
semantic network
frame
knowledge reasoning
definition
structure
Analyze and synthesize multiple judgments to make new judgments
process
knowledge processing
Main forms of mental processes
syllogism
Linear (linear syllogism)
condition
Probability
reasoning method
reasoning logic
interpretation
syllogism
induction
Assumptions and proofs
knowledge certainty
Sure
uncertain
Machine implementation
Reasoning engine (program that implements reasoning)
Reasoning control strategy
Classification
reasoning strategies
search strategy
Deterministic knowledge representation method
predicate logic notation
Basics of Logic
proposition
Declarative sentences (assertions) with true or false meanings
truth value (the meaning of a proposition)
T/F
domain of discourse
The set composed of all the objects in question also becomes the individual domain
predicate
Used to represent propositions in predicate logic, such as P(x1,x2,...,xn). P:D^n->{T,F}. where D^n={(x1,x2,...,xn)|x1,x2,xn∈D)
It is a mapping from D to {T, F}. The true value is T or F and can exist independently.
function
f:D^n->D. where D^n={(x1,x2,...,xn)|x1,x2,xn∈D)
D to D mapping, the value of the function is the element in D, which can only exist as an individual of the predicate
conjunction
┐, ∧, ∨, →, ↔
quantifier
∀, ∃
scope of quantifier
Refers to a single predicate or formula after a quantifier
constraint variables
The variable with the same name as the quantifier in the domain
free variable
Unconstrained variables
Classic example of predicate logic representation
condition part
action part
Delete table
Add table
Features
advantage
Natural, clear, precise, flexible, rigorous, modular and easy to implement
shortcoming
Poor knowledge representation ability
can only represent deterministic knowledge
Knowledge base management difficulties
Lack of knowledge of organizing principles
There are too many combinations
Can only blindly reason
System efficiency is low
Separation of reasoning calculus and knowledge meaning
production notation
basic method
fact
concept
A fact is a statement that asserts the value of a linguistic variable or a relationship between multiple linguistic variables.
display method
(object, attribute, value)
(relationship, object 1, object 2)
rule
form
P->Q
IF A AND B THEN C
characteristic
advantage
natural, modular, effective
Low efficiency and inconvenient to express structural knowledge
semantic network representation
definition
Entities and semantic relationships to express knowledge in directed graphs
composition
node
arc
semantic unit
(node 1, arc, node 2)
Basic network element
basic semantic relations
Instance relationship: ISA
Meaning: is a
Embodiment: Concrete and Abstract
Classification relationship (generalization relationship): AKO
Meaning: a
Embodiment: subclasses and superclasses
Membership: A-Member-of
Meaning: to be a member
Embodiment: individual and collective
Attribute relationship: Have
Meaning: Yes
Embodiment: having an attribute
Attribute relationship: Can
Meaning: can
Manifestation: being able to do something
Inclusion relationship (clustering relationship): Part-of
Meaning: part of
Embodiment: part and whole
Note: There is no inheritance of properties
Time relationship: Before/after
Meaning: before/after
Positional relationship: Locat-on/Locat-under/Locat-at
Meaning: above/under/on…
Similar relationship: Similar-to/Near-to
Meaning: similar/close
reasoning
inherit
Abstraction passed to instance
Create a node table and find ISA, AKO, A-Member-of
match
Find matches
Construct a fragment
characteristic
advantage
Structural, associative, natural
shortcoming
non-strictness
Complexity
frame notation
frame theory
frame
As long as people add new things to the framework, they can form a specific entity
instance frame
For a framework, after people fill in the details, they get concrete examples.
frame system
In frame theory, frame is the basic unit of knowledge
Frame structure and frame representation
Frame<frame name>
Slot name 1: Side name 1 value 1, value 2,...
frame system
portrait
AKO and ISA
Horizontal
Characteristics of frame representation
advantage
Structural, deep, inherited, natural, able to express cause and effect
shortcoming
Lack of formal theory of framework, lack of process knowledge, difficulty in ensuring clarity
Deterministic knowledge reasoning method
production reasoning
basic structure
comprehensive database
Store various information for the reasoning process
initial state of the problem
input facts
Intermediate and final conclusions
as a basis for selection of reasoning processes
Rule base (knowledge base)
effect
Stores all the rules needed for reasoning
It is the basis for production systems to be able to reason.
Require
Complete, consistent, accurate, flexible, organizeable
Control system (inference engine)
effect
Control system operation
Task
Select match
Select rules from the rule base according to a certain strategy and match them with known facts in the comprehensive database
conflict resolution
For successfully matched rules, execute them according to a certain strategy
perform operations
Add the conclusions drawn to the comprehensive database and continue execution if there are other operations.
Termination reasoning
Check if the synthetic database contains the target and stop inference if so
Lu Jin explains
Remember the sequence of rules applied and give an explanation path for the problem
reasoning method
forward reasoning
reverse reasoning
hybrid reasoning
natural deductive reasoning
Logical basis
Equivalence
Yongzhen style
replace
unity
method
Push hard
inductive deductive reasoning
The logical basis of inductive and deductive reasoning
true and false
eternal authenticity
Any one is satisfied
Satisfiability (compatibility)
There is at least one individual whose value is true
Permanent falsity (incompatibility)
Not satisfied
paradigm
toe-beam paradigm
All quantifiers appear non-negatively at the front of the formula and govern the entire formula.
Skolem paradigm
Based on the front bundle paradigm, all existential quantifiers are in front of the universal quantifier.
Clause sets and their simplification
clause
Word
Atomic predicate formula and its negation
clause
The disjunction of any literal becomes a clause
empty clause
clause without any text
permanent vacation
Remember as □ or NIL
clause set
In predicate logic, any predicate formula can be turned into a corresponding set of clauses by applying equivalence relations and inference rules.
Simplification of clause sets
Eliminate "→" and "↔" in predicate formulas
P→Q can be written as ┐P∨Q
P↔Q can be written as (P∧Q)∨(┐P∧┐Q)
Reduce the scope of negation symbols so that they only apply to one predicate
┐(┐P) can be written as P
┐(P∧Q) can be written as (┐P)∨(┐Q)
┐(P∨Q) can be written as (┐P)∧(┐Q)
┐(∀x)P(x) can be written as (∃x)┐(P(x))
┐(∃x)P(x) can be written as (∀x)┐(P(x))
Standardize the variables
Rename the quantifier so that the variables constrained by different quantifiers have different names.
toe-in paradigm
Move all quantifiers to the left of the formula, being careful not to change their relative order when moving.
eliminate existential quantifier
If there is a quantifier that does not appear in the scope of the universal quantifier (that is, there is no global quantifier on its left)
Replace the variable of the constraint with a new individual constant
For example, (∃x)P(x) can be written as P(y)
If there is a quantifier in the scope of the universal quantifier (that is, there is no global quantifier on its left)
Replace the y argument with the Skolem function
If y is within the scope of x, then y can be written as f(x)
into Skolem standard form
Reduce the jurisdiction of ∨ so that ∨ only acts on one predicate
P∨(Q∧R) can be written as (P∨Q)∧P(P∨R)
eliminate universal quantifier
Because all variables are constrained by universal quantifiers and global quantifiers have nothing to do with order.
can be omitted directly
Eliminate the conjunction ∧
Turn the predicate formula into a set of clauses
Change variable name
The same argument name does not appear in any two clauses
characteristic
Not unique due to Skolemization, but does not affect satisfiability
The necessary and sufficient condition that the predicate formula F cannot be satisfied is that the set of clauses cannot be satisfied
Robinson's principle of reduction
Basic idea
There is a conjunctive ∧ relationship between the clauses. Therefore, as long as one clause is not satisfied, the entire set of clauses is not satisfied.
If a clause set contains an empty clause, then the clause set cannot satisfy
First, deny the question you want to prove and add it to the set of clauses. Verify that the clause set has an empty clause. If there is an empty clause, it means that the negation of the question is false, otherwise the reduction continues. It's true if it's true no matter how you boil it down.
propositional logic
reductive reasoning
complementary text
If P is an atomic predicate formula, then P and ┐P are complementary
come down to
C1C2 is a clause (only disjunction ∨). C1 has L1 and C2 has L2. If L1 and L2 are complementary, L1L2 is eliminated. Press the disjunction ∨ on the remaining part to get the new clause C12. C1 and C2 are parent clauses of C12
characteristic
The results are the same, but the process is not unique
in conclusion
C12 is the logical conclusion of C1 and C2
If C1 and C2 are true, then C12 must be true
After C12 replaces C1 and C2, the unsatisfiability of the new clause set S1 can be derived from the unsatisfiability of S
After C12 joins C1 and C2, the unsatisfiability of the new clause set S1 can mutually infer the unsatisfiability of S.
predicate logic
boil down to logic
When there are no public variables, a unification operation is performed, recorded as {a/y}
When there are public variables, replace them first and then combine them into one.
Two pairs cannot be eliminated at the same time
If you go around in circles or have no solution, check whether there is a problem with the process.
If there are internal variables that can be unified, they should be unified internally first.
Methods of deductive reasoning
principle
Machine reasoning method based on Robinson reduction principle
process
negative goal formula
Put your goals into a formula set
Convert the set of formulas into a set of clauses
Summarize a set of clauses
A simple example of deterministic knowledge system
Uncertainty Reasoning
Basic concepts of uncertainty reasoning
The meaning of uncertainty reasoning
Starting from uncertain initial evidence, using uncertain knowledge to derive uncertain but reasonable or basically reasonable conclusions
The scope of application of uncertainty reasoning
Incomplete and inaccurate knowledge
vague description
Multiple reasons lead to the same conclusion
The result is not unique
Basic Issues in Uncertainty Reasoning
Uncertainty representation of knowledge
Considerations
Problem description ability, calculation of uncertainty in reasoning
meaning
The degree of certainty of knowledge, or static strength
display method
Probability[0,1]
Credibility[-1,1]
Uncertainty Expression of Evidence
Type of evidence
evidence organization
basic evidence
Combining evidence
Disjunction or union. Based on basic evidence, there are max-min methods, probabilistic methods, bounded methods, etc.
Source of evidence
initial, middle
display method
Probability, credibility, fuzzy sets
Uncertain matching
meaning
Uncertain preconditions match uncertain facts
Calculation method
Design an algorithm to calculate the degree of similarity and give a limit of similarity to see if it falls within the limit.
Uncertain Update
How to use uncertainty in evidence to update uncertainty in conclusions
Different methods handle it differently
If the uncertainty of intermediate conclusions is conveyed to the final revelation
Put the current conclusion and its uncertainty into the database as a new conclusion and pass it
Synthesis of Uncertain Conclusions
Multiple different knowledges lead to the same conclusion, but with different credibility
Types of Uncertainty Reasoning
Numerical Methods
based on probability
credibility approach
Subjective Bayes
evidence theory
probabilistic reasoning
Fuzzy reasoning
non-numerical methods
credibility reasoning
The concept of credibility
Credibility is the degree to which people believe that a certain thing or phenomenon is true, and it has a certain degree of subjectivity
credibility inference model
Rule representation
IF E THEN H(CF(H,E))
definition
E is the premise, H is the conclusion, and CF is credibility.
CF(H,E)
The value range is [-1,1]
CF(H,E)=MB(H,E)-MD(H,E)
definition
MB is trust growth
MD is the growth rate of distrust
in conclusion
Mutually exclusive, one of the two must be 0
The value range is [-1,1]
Can push back and forth
MB(H)=MD(┐H)
MB(H) MB(┐H)=0
Credibility is not probability and does not satisfy P(H) P(┐H)=1
calculate
Calculate the confidence of a single piece of evidence in a conclusion
Update formula
CF(H)=CF(H,E)*max(0,CF(E))
Do not consider the impact of CF being false on H
Combining evidence
Conjunction and: take the minimum value
Disjunction or: take the maximum value
Calculate the confidence of a single piece of evidence in a conclusion
Synthetic formula
If the division cannot be completed, round to two decimal places.
application
Hematology diagnosis expert system MYCIN
Subjective Bayesian Reasoning
Basics of Probability Theory
Total probability formula, Bayesian probability
inference model
main method
Update the posterior probability of H
Rule representation
IF E THEN (LS,LN) H
definition
(LS,LN)
are used to represent the knowledge intensity of the knowledge
LS
sufficiency measure
LN
necessity measure
O(X)
probability function
P(x)=0↔O(x)=0
P(x)=1↔O(x)=∞
O(X)=P(X)/(1-P(X))
P(X)=O(X)/(1 O(X))
in conclusion
O(H|E)=LS*O(H)
O(H|┐E)=LN*O(H)
nature
LS
LS>1
E supports H. The larger LS is, the more fully E supports H.
When LS approaches ∞, P(H|E) approaches 1, which means that E will cause H to be true.
LS=1
E has no effect on H
LS<1
E does not support H. The smaller LS is, the more fully E does not support H.
When LS=0, P(H|E)=0, which means that E will cause H to be false
LN
Replace the E in LS with ┐E
calculate
Calculate the confidence of a single piece of evidence in a conclusion
Update formula
When E is definitely true
When E is definitely false
When E has nothing to do with S
Other cases
Combining evidence
Look at P(E|S)
Conjunction and: Take the minimum value and directly discard the larger piece of evidence.
Disjunctive or: take the maximum value and directly discard the smaller piece of evidence
Calculate the confidence of multiple pieces of evidence on a conclusion
In the middle of calculation, the composition is synthesized and can be inherited later.
application
Geological and mineral exploration expert system PROSPECTOR
Fuzzy reasoning
fuzzy set
concept
μF is the membership function, F is the fuzzy set of U
The larger μ(u) is, the higher the degree of membership is.
Fuzzy sets and their membership functions are not equivalent
express
Discrete
Notice
/ is the delimiter, not the division sign
It's a connector, not a plus sign
When the membership degree is 0, it can be omitted, but it is not recommended.
example
F=1/20 0.8/30
Continuous
example
General representation method
Notice
∫ is a relational symbol, not an integral symbol in mathematics.
Operation
F⊆G
For any u belonging to U, there is μF(u)<=μF(u)
F∪G
For each u, take the maximum value of the two
F∩G
For each u, take the minimum value of the two
Subtopic 2
┐F (complement)
For each u, take 1-μF(u)
fuzzy concept matching
semantic distance
definition
example
effect
suitability
definition
1-d(F,G)
A threshold can be given to determine whether two concepts match
closeness
Note that the outer product must supplement the omitted elements with a membership of 0.
fuzzy relationship
definition
Similar to Cartesian product
The construction of fuzzy relationships
definition
Similar to the Cartesian product, first take the smaller value and then take the larger value
example
U is row, V is column
Synthesis of Fuzzy Relationships
definition
Similar to matrix multiplication, when calculating each position, "take the small one first and then the large one"
example
blur transformation
The synthesis of operations and fuzzy relations
Fuzzy inference method
According to the given reasoning mode, it is realized through the synthesis of fuzzy sets and fuzzy relations.
Fuzzy knowledge representation
rule
IF x is F THEN y is G
evidence
x is F`
Fuzzy inference method
fuzzy hypothetical reasoning
definition
example
fuzzy resistant reasoning
definition
example
hypothetical syllogism
definition
example
Intelligent search technology
Search overview
meaning
search
Use knowledge to construct a reasoning route with the least cost
Smart search
An algorithm that uses intermediate information obtained during the search process to guide the search in the optimal direction.
type
Based on space search
state space
A algorithm
A* algorithm
AND/OR TREE
problem reduction method
game tree
Maximum/minimum algorithm
α-β pruning
Based on random algorithm
Evolution mechanism
genetic algorithm
immune optimization
immune algorithm
Population optimization
Ant Colony Algorithm
Granular Swarm Algorithm
statistical model
Model Carlo algorithm
Other methods
Hill Climbing Search Algorithm
simulated annealing algorithm
State space problem solving method
State space problem representation
state
A data structure that represents the problem status of each step in the problem solving process, which can be represented by a vector
operate
Also called an operator, it transforms one state into another state and describes the relationship between states.
state space
Used to describe all states of a problem and the relationships between states, which can be represented by a triple (S, F, G)
S
initial state
F
Operation set
G
target state
can be represented by a directed graph
Node represents state
Edges represent operations
Problem solving process
Choose the appropriate status and action
Starting from some initial state, build a sequence of operations using one operation at a time until the goal state is reached
The sequence of operators used from the initial state to the goal state is a solution
Problem specification solving method
Basic idea
Decompose or equivalently transform the problem into a series of simple sub-problems
The process of performing a search on an AND-OR graph with the goal of indicating that the starting node is solvable. That is, the search is not to find a path to the target node, but to find a solution tree.
display method
AND/OR TREE
definition
end node
Node without child nodes
Termination node
original problem
Questions that can be answered directly (recursive exit)
The node corresponding to the original problem
Solvable nodes
Any terminal node is a solvable node
If it is an 'OR' node, when at least one of its child nodes is a resolvable point, it is a resolvable point.
If it is an 'AND' node, when all the child nodes are resolvable points, it is a resolvable point.
solution tree
It consists of solvable borrowing points, and from the solvable nodes, it can be deduced that the initial node is a solvable borrowing point. Such a subtree is a solution tree.
Blind strategies for searching
depth and breadth
Heuristic search of state space
inspiring information
Information that can guide your search
valuation function
Estimate node importance
f(n)=g(n) h(n)
g(n)
The actual cost from the initial node S0 to node n
h(n)
The estimated cost of the optimal path from n to target node Sn
A algorithm
Select an expansion with the smallest estimate from all nodes in the Open table
A* algorithm
Make the following restrictions on g(n) and h(n) in algorithm A
g(n) is an estimate of the minimum cost g*(n), and g(n)>0
h(n) is the lower bound of the minimum cost h*(n), that is, h(n)<=h*(n)
Heuristic search with/or trees
The cost of solving the tree
If n is the terminal node, the cost h(n)=0
If n is an OR node
Cost h(n)=(cost of child node cost of node n to child node) minimum value
If n is an AND node
and consideration method
Cost h(n)=(cost of child node cost of node n to child node) and then sum
maximum cost method
Cost h(n)=(cost of child node cost of node n to child node) maximum value
If n is an end node, but not a terminal node
Cost h(n)=∞
hope tree
The tree that is most likely to become the optimal solution tree during the search process
Generally, two layers are expanded at a time.
Heuristic search for game trees
game
Features
Your own turn is "OR", and the opponent's turn is "AND"
Alternate with or nodes
minimax process
Generate a partial game tree and evaluate the leaf nodes
leaf node
A positive value is taken when it is beneficial to us, and a negative value is taken when it is beneficial to the other party.
non-leaf node
Push upward from leaf nodes
For our node, we select the node with the largest valuation every time, so the value of our node should be the maximum value of the successor node.
For the opponent node, the node with the smallest valuation is selected every time, so the value of our node should be the minimum value of the successor node.
α-β pruning
Edges generate nodes and edges evaluate nodes.
Our node->α value: the maximum value of the current child node
Opposite node->β value: the minimum value of the current child node
Default is 'I' go first
pruning method
alpha pruning
If the beta value of any opponent node is less than or equal to the alpha value of the predecessor node, the search will stop.
The root node of the pruned subtree is at the α level, which is α pruning.
β pruning
If the α value of any of our nodes is greater than or equal to the β value of the predecessor node, the search will stop.
The root node of the pruned subtree is at the β level, which is β pruning.
genetic algorithm
definition
process
coding
type
binary encoding
definition
Transform the structure of the original problem into the bit string structure of the chromosome
operate
First determine the length l, which is related to the domain of the variable and the calculation accuracy
shortcoming
In binary, 7 and 8 are very similar, but there is a big difference from 0111 to 1000 (Hamming Cliff)
Gray code
definition
An improvement to binary encoding that requires that the encoding of two consecutive integers can only differ by one code bit.
operate
real encoding
Features
floating point
Scope of application
High precision, multi-dimensional
Character Encoding
fitness function
Commonly used fitness functions
original fitness function
advantage
Directly reflects the original goal
shortcoming
Negative numbers may occur
standard fitness function
extremely small
great
accelerated transformation
Basic operations
Select action
Ratio selection
Roulette
The greater the fitness value, the greater the possibility of being selected.
Breeding pond
sort selection
competitive options
Tournament selection
crossover operation
definition
partial genetic recombination
type
binary crossover
single point crossover
No change before a certain point, interchange after a certain point
Two points cross
An exchange between two points, unchanged before and after
multi-point crossover
evenly cross
substantial crossover
mutation operation
type
binary mutation
Changes between 0↔1
substantial variation
In the process of biological evolution, selection works through inheritance and mutation, and at the same time, mutation and inheritance develop in the direction of adapting to the environment.
machine learning
Machine Learning Overview
learning concept
psychological perspective on learning
Based on brain science and cognitive science
Classification
connectionist perspective
The essence of learning is the formation of connections
epistemological perspective
The essence of learning is the change of cognitive structure in the learner's mind
Change
Behavior
behavioral potential
core
Marked by changes in behavior and behavioral potential
Behavioral changes caused by experience
Behavioral changes last longer
An artificial intelligence perspective on learning
acquisition of knowledge
core
improvement of abilities
result
general explanation
Machine learning concepts
What is machine learning
Machines simulate human learning
Main research content
cognitive simulation
theoretical analysis
task-oriented research
machine learning system
Have an appropriate learning environment
Have certain learning ability
Ability to use learned knowledge to solve problems
Improve your own performance through learning
Types of machine learning
Is there guidance from a mentor?
supervised learning
Classified learning style
unsupervised learning
generative learning approach
learning strategies
Memory learning, teaching learning, deductive learning, inductive learning
Application areas
Expert-identical learning, robot learning, natural language understanding learning
Basic model of machine learning system
environment
The learning system perceives a collection of external information, which is also an external source.
learning session
Organize, analyze and form information provided by the environment and put it into the knowledge base
knowledge base
Store processed information
Execution link
Perform tasks based on the knowledge base
memory learning
Memoized search
Trade space for time
storage structure
environmental stability
Memory versus Computation Tradeoff
Simon's Checkers Program
Learn by example
type
Example source
Teachers, learners themselves, external environment
Example type
Use only positive examples
Use positive and negative examples
Model
sample space
collection of examples
induction process
The inductive process of abstracting general knowledge from examples
regular space
collection of various laws
Verification process
Select new instances from the example space to further verify and modify the rules just summarized.
Positive examples produce concepts, while counterexamples do not. Positive examples expand the extension, and counterexamples narrow the extension.
inductive method
constant to variable
remove conditions
Add options (∨ operation)
Curve fitting (least squares method)
Decision tree learning
The concept of decision tree
In a decision tree, there is a conjunctive relationship between all attributes on the same path.
ID3 algorithm
concept
A learning algorithm that uses the decreasing speed of information entropy as the attribute selection criterion
Split subtrees based on the principle of maximizing the decrease in information entropy, and gradually construct a decision tree.
information entropy
definition
Information entropy is a measure of the overall uncertainty of the information source. The smaller its value, the smaller the uncertainty of the information source.
Entropy quantification formula
Probability and information entropy are two sides of the same coin of random events
Probability
How likely is something to happen?
Characterizing certainty
information entropy
How many possibilities does something have?
Characterizing uncertainty
calculate
Evenly distributed
An event has m equally probable situations, then its entropy is n=log(m)
base
2
Bits
Common, default
e
Nat
Differential calculation
10
Bell
linguistics
general distribution
definition
understand
joint entropy
information gain
definition
A measure of the difference between two quantities of information
weighted information entropy
formula
step
(4)
Calculate information entropy E(S,X)
Note: S is a node. If it is a child node, all S under it must be replaced by Si. i is the number of the child node.
For each attribute xi separately
Calculate its corresponding weighted information entropy E(S,xi)
First write down what elements and proportions there are in St, calculate the information entropy E(St,X) of St, and repeat this process until all calculations are completed.
Then bring in the formula to calculate E((S,X),xi)
Calculate its corresponding information gain G(S,Xi)
Extend the attribute with the greatest information gain
Artificial Neural Network Connection Learning
Overview
Artificial neural networks
concept
Artificial neurons are connected topologically
abbreviation
ANN
structure
input, calculation, output
Connected learning
Complete the correction and stabilization of synaptic connection weights
Classification
Shallow layer
sensor
BP network
Hopfield network
Deep
Biological mechanism of artificial neural network
structure
dendrites
enter
cell body
calculate
axon
output
type
Classification according to the number of protrusions
Classification according to the electrophysiological properties of neurons
Classification according to the function of neurons
Structure and model of artificial neurons and artificial neural networks
Structure and model
artificial neurons
structure
MP model
type
threshold type
Piecewise linear strongly saturated type
S type
Sub-threshold cumulative type
The neuron model can be viewed as a simple classifier
Artificial neural networks
concept
A network formed by the interconnection of artificial neurons
Classification
Features
massively parallel processing
Distributed storage of information
Have learning ability
The interconnection (topology) structure of artificial neural networks
feedforward network
concept
Contains only forward joins
single layer feedforward network
It only contains input layer and output layer, and only the neurons of the output layer are computable nodes.
multilayer feedforward network
In addition to the input layer and output layer, it also contains at least one or more hidden layers.
BP network
feedback network
concept
Can contain feedback connections and loop connections
single layer feedback network
Feedback network without hidden layers
multi-layer feedback network
Feedback network with hidden layer
Hopfield network
Stability is described by the energy function
Shallow model of artificial neural network
perceptron model
Purpose
Classify external input
If the external input is linearly separable (satisfying Σ(Wij*xi-θj)=0), then it can be divided into two categories
The network topology of a single-layer perceptron is a single-layer feedforward network.
Single layer perceptron solves linearly separable problems
A single-layer perceptron contains an input layer and an output layer
perceptron
The perceptron actually classifies different points by constructing a hyperplane
The concept of learning was introduced for the first time in the perceptron, which simulated the learning function of the human brain to a certain extent.
The biggest difference between the perceptron and the neuron model is that the perceptron model can learn from training samples