Hard Optimization Problems: Practical Approach DORIT RON Tel Ziskind room #303
Course outline 1 st lecture: Introduction and motivation
INTRODUCTION What is an optimization problem?
An optimization problem consist of: Variables: Energy functional to be minimized/maximized: min / max
Unconstrained minimization Find the global minimum
An optimization problem consist of: Variables: Energy functional to minimized/maximized: min / max Possibly subject to: Equality constraints: Inequality constraints:
Constrained minimization subject to
INTRODUCTION What is an optimization problem? Examples
Example 1: 2D Ising spins Discrete (combinatorial) optimization min - i,j> s i s j s i = { +1, -1}
3D Ising model Each spin represents a tiny magnet The spins tend to align below a certain T c Ferromagnet – Iron at room temperature magnet T c non-magnet ----| | |--> T room temp ferromagnetism melting 770 o C 1538 o C At T=0 the system settles at its ground states
Example 2: 1D graph ordering Given a graph G=(V, ), find a permutation of the vertices that minimizes E( )= i j w i j | (i) - (j) | p where i, j are in V and w i j is the edge weight between i and j (w i j =0 if ij is not in ) p=1 : Linear arrangement p=2 : Quadratic energy p= : The Bandwidth
i j Minimum Linear Arrangement Problem
i j
Minimum Linear Arrangement Problem i j
Minimum Linear Arrangement Problem i j
Minimum Linear Arrangement Problem i j
Minimum Linear Arrangement Problem i j
General Linear Arrangement Problem Variable nodes sizes E(x)= i j w i j | x i -x j | p x i = v i /2 + k: k)< i) v k i j xixi xjxj
Other graph ordering problems Minimize various functionals: envelope size, cutwidth, profile of graph, etc. Traveling salesman problem – TSP
The Traveling Salesman Problem
Other graph ordering problems Minimize various functionals: envelope size, cutwidth, profile of graph, etc. Traveling salesman problem – TSP Graph bisectioning Graph partitioning Graph coloring Graph drawing
Drawing Graphs
Example 3: 2D circuit placement Bottleneck in the microchip industry Given a hypergraph, find the discrete placement of each module (gate) while minimizing the wirelength
The hypergraph for a microchip
Placement on a grid of pins
Routing over the placement
Example 3: 2D circuit placement Bottleneck in the microchip industry Given a hypergraph, find the discrete placement of each module (gate) while minimizing the wirelength No overlap is allowed No overflow is allowed Critical paths must be shorter Leave white space for routing Typical IBM chip ~270 meters on 1cm 2
Place and route
Exponential growth of transistors for Intel processors
INTRODUCTION What is an optimization problem? Examples Summary of difficulties
Difficulties: Many variables: 10 6, 10 7 … Many constraints: 10 6, 10 7 … Multitude of local optima Discrete nature Conflicting objectives Reasonable running time
INTRODUCTION What is an optimization problem? Examples Summary of difficulties Is the global optimum really needed / obtainable?
PEKO=PLACEMENT EXAMPLE WITH KNOWN OPTIMUM Place the nodes – this is the solution Create the net list locally and compactly The optimum wire length – the sum of all the edges between the nodes, is known and can be proven to be minimal
SOLUTION QUALITY
INTRODUCTION What is an optimization problem? Examples Summary of difficulties Is the global optimum really needed / obtainable? What is expected of a “good approximate” solution?
“Good approximate” solution As optimal as possible: high quality solution Achievable in linear time Scalable in the problem size
RUNTIME
Reality Check Rigorous Optimization Theorems LIMITED Industrial Need for FAST & GOOD NP-Complete Intractable Problems HEURISTICS
INTRODUCTION What is an optimization problem? Examples Summary of difficulties Is the global optimum really needed / obtainable? What is expected of a “good approximate” solution? Multilevel philosophy
MULTILEVEL APPROACH PARTIAL DIFFERENTIAL EQUATIONS (Achi Brandt since the early 70’s) STATISTICAL PHYSICS CHEMISTRY IMAGE SEGMENTATION TOMOGRAPHY GRAPH OPTIMIZATION PROBLEMS
SOLUTION QUALITY
ORIGINAL PICTURE
ORIGINAL FENG SHUI (1) FENG SHUI (2) mPL KRAFTWERK CAPO DRAGON OURS
OUR PLACEMENT
Course outline 1 st lecture: Introduction and motivation 2 nd – 4 th : Local processing (relaxation) Quadratic minimization, Newton’s method, Steepest descent, Line search, Lagrange multipliers, Active set approach, Linear programming 4 th – 5 th : Global approaches Simulated annealing, Genetic algorithms, Spectral method 6 th : Classical geometric multigrid 7 th : Algebraic multilevel 8 th : Graph coarsening 9 th – 12 th : Advanced multilevel topics