Application of Optimization Methods and Edge AIhliangzhao.me › slides › Optimization_and_Edge_AI.pdfApplication of Optimization Methods and Edge AI Hai-Liang Zhao [email protected]

Application of Optimization Methods Edge AI

Application of Optimization Methodsand Edge AI

Hai-Liang [email protected]

November 17, 2018

This slide can be downloaded at Link .

https://narcissushliangzhao.github.io/Curriculum-Vitae/slides/optimization_edge_AI.pdf


Outline

1 Application of Optimization Methods. Existing Methods and Their Applications. How to Design Novel Models with Methods EmbeddedNaturally?

2 Edge AI. Existing Paradigms. Preparation for Designing Egent


Outline




Outline




. Existing Methods and Their Applications

Regular Optimization Methods

1 Evolutionary Algorithms

2 Lyapunov Optimization

3 Stochastic Programming

4 Game Theory

5 Traditional Machine Learning Methods

6 Various Programming Methods

7 Deep Reinforcement Learning

8 ...



Evolutionary Algorithms

1 Swarm Intelligence

2 Tabu Search

3 Simulated Annealing

4 Artificial Neural Networks5 Population-based Algorithms

1 genetic algorithm2 particle swarm optimization3 negative selection algorithm4 learning-teaching-based optimization

6 Too many of them ...

Many works on Service Composition contributed by Prof.Shuiguang Deng are solved by Evolutionary Algorithmsbecause they are method-free.



Lyapunov Optimization

Standard Lyapunov Optimization is a trump card forstochastic optimization problems.

1 Virtual Queues

2 Drift-Plus-Penalty Expression

3 Approximate Scheduling4 Performance Analysis

1 average penalty analysis2 average queue size analysis

5 Trade-off by Tuning V

A brief introdution for researchers can be found at Link .

https://narcissushliangzhao.github.io/Curriculum-Vitae/monographs/Lyapunov_optimization.pdf



Applications of Lyapunov Optimization

Yuyi Mao’s papers are inundated with this kind ofmethods.

1 Match with Lyapunov Optimization Methods1 Construct Virtual Queues for Constraints2 Replace the Original Problem with a Deterministic one3 Solve the Approximate-Convex Problem with Ingenious

Mathematic Tricks

2 Utilize Lagrange Methods and KKT Conditions

3 Performence Analysis: O(V ), O( 1V )

Apperently Yuyi Mao acquires prociency in Michael. J. Neely’sbook: Stochastic Network Optimization withApplication to Communication and Queueing Systems



Extensions on Lyapunov Optimization

1 Extensions to Variable Frame Length Systems (DynamicOptimization and Learning for Renewal Systems)

2 Combination with Lagrange Multipliers

3 Network Utility Maximization over Partially ObservableMarkovian Channels

4 Under Non-Convex Problems (Greedy primal-dualalgorithm)



My work

P : maxIt

limT→+∞

1

T

T−1∑t=0

E[∑i∈NUi(It

i,:)

],

where Ui(Iti,:) is defined as

Ui(Iti,:) ,

∑j∈M

rti,jIti,j − φti ·

[ ∑j∈M

εti,jIti,j − ψ

safei , 0

]+.

s.t. It ∈ {0, 1}N×M , t ∈ T ,∑i∈N I

ti,j ≤ Nmax

j , j ∈M, t ∈ T ,∑j∈M εti,jI

ti,j ≤ ψt

i , i ∈ N , t ∈ T .



Stochatsic Programming

Two-stage or Multi-stage ⇓1 Scenario construction

2 Monte Carlo techniques (SAA method)

3 Evaluation Candidate Solutions (measure the optimalitygap between the optimal value and the estimated value)



My work

P : minΘ′∈Q

EP [C(Θ′,D)] ,∑i∈N

(Ii · EP [clocal?i ]

+∑

j∈Mk

Oij · EP [ctx?ij ]

+∑

j∈Mk

∑j′∈Mkj

\{j}

Rkjij′ · EP [creloc?

ijj′]

+∑

j∈Mk

∑j′∈Mk

Rkjij′ · EP [cserver?ij ]

),

with 8 Constraints.



Deep Reinforcement Learning

Model:



Deep Reinforcement Learning

Method:



My work

P : Φ? = argmaxΦ

EΦ

[(1− γ)

∞∑t=1

γt−1r(xt,Φ)|x1 = x

],∀x ∈ X ,

where r(xt,Φ) is defined by

r(xt,Φ) =∑i∈N

∆ti − % ·

(φ ·∑j∈M

(1{Its = j}

+ 1{Itd = 1})

+ ζ ·∑i∈N

1{Qti ≤ 0}

+ ξ ·∑i∈N

Eti (s) ·

(1{s /∈ Sti ∧ Et

i (s) 6= 0})),

where ∆ti is defined as

∆ti =

∑s∈{s′∈S|At

i,s′=1}

(1{Its = j?} ·

(bti,c(s)− bti,e(s, j?)

)).


. How to Design Novel Models with Methods Embedded Naturally?

My Questions

How to Design Novel Models with MethodsEmbedded Naturally?D Fullfill understanding on Convex Optimization?D Stop Compromising of system model for fancy

mathematical derivation?D Find Stream and Tide?


Outline




. Existing Paradigms

Existing Paradigms

. Use Machine Learning Methods to solve traditionalresource management or latency minimizationproblems. Design new structure with heterogeous edge sites forAI-enabled apps

1 Edge Federated Learning

Federated Learning

Federated Learning features distributed learning † at edgedevices and model-update aggregation‡ at an edge server.

2 Learning-driven Communication


. Existing Paradigms

Learning-driven Communication

Conventional philosophy in traditional wireless communication

The traditional design objectives of wireless communications,i.e., coomunication reliablity and data-ratemaximization , do not directly match that of edge learning.

ww� A conceptual change

Learning-driven communication

The coupling between communication and learning in edgelearning systems should be exploited.

A more detailed slide on Learning-driven Communication canbe found at Link

https://narcissushliangzhao.github.io/Curriculum-Vitae/slides/learning_driven_commun.pdf


. Preparation for Designing Egent

Preparation for designing Egent

Network models compactable for Egent

Knowledge on Communication in this paper greatly enlightensme the desgin compactable network/communication models,which is a bottomed layer of Egent.

Noise in training-data transmission maybe is not thatimportant.

Long-term observations and collected data on user profilescan be utilized for joint resource manegment andmodel-training.

Collaboration between cloud and edge learning can bein-depth studied.

Mobility management of users not only effects theoffloading decisions, but also incurs frequent handoversamong edge servers (not service migration).

Documents

Application of Optimization Methods and Edge AIhliangzhao.me › slides › Optimization_and_Edge_AI.pdfApplication of Optimization Methods and Edge AI Hai-Liang Zhao [email protected]