Download presentation
Presentation is loading. Please wait.
1
Some issues of design philosophy…
Basic Concepts Some issues of design philosophy…
2
Memory-Processor Link and the von Neumann Bottleneck
Processor = CPU Memory = hierarchy of caches, main memory, and bulk storage Connection = address, control, data busses Source of contention = bus traffic during the FETCH/EXECUTE cycle 8/23/2019 Copyright G.A. Tagliarini, PhD
3
Memory-Processor Link
Biological systems Memory unit = processor Large numbers of independent but interacting processors (circa 1010 neurons) Small computational complexity Massive connectivity (circa 1012 synapses) Connections and state represent solutions 8/23/2019 Copyright G.A. Tagliarini, PhD
4
Common Characteristics of BIP
Parallel distributed processing Rummelhart, McClelland and Hinton Stigmergy Feedback Negative Positive 8/23/2019 Copyright G.A. Tagliarini, PhD
5
Copyright G.A. Tagliarini, PhD
Emergent Behavior Adaptation Learning Supervised – a teacher instructs using a set of examples and the student generalizes to produce a representation Unsupervised – examples are presented and a rule for forming group representations may be given Evolution Fitness/selective pressure Crossover/recombination Mutation 8/23/2019 Copyright G.A. Tagliarini, PhD
6
Emergent Behavior (continued)
Self-organization Cooperation Coordination Order arises without external intervention Contrast with Templates Supervisors Instructions/blueprints Key question: How to formulate behaviors that give rise to self-organization? 8/23/2019 Copyright G.A. Tagliarini, PhD
7
Example: How to Get Ants to Cooperate when Foraging
Suppose ants have been endowed with search behaviors and individual carrying capacity? Allow the reward structure to be changeable What behaviors or conditions would elicit cooperation? 8/23/2019 Copyright G.A. Tagliarini, PhD
8
A Search Problem: Satisfiability
Given A set of m logic variables {p1, p2, …, pm} A collection of n disjunctive clauses, e.g., ci = (p1 + p2’ + p5) Place the clauses in conjunctive normal form (CNF) to create c = c1 c2 …ci …cn Question Is there an assignment of truth values that satisfies c? “…satisfies c…” means “…for which c is true…” See further: 8/23/2019 Copyright G.A. Tagliarini, PhD
9
An Optimization Problem: Maximum Satisfiability
Given A set of m logic variables {p1, p2, …, pm} A collection of n disjunctive clauses, e.g., ci = (p1 + p2’ + p5) Place the clauses in conjunctive normal form (CNF) to create c = c1 c2 …ci …cn Question What is the assignment of truth values that maximizes the number of clauses satisfied? 8/23/2019 Copyright G.A. Tagliarini, PhD
10
An Optimization Problem: Weighted Maximum Satisfiability
Given A set of m logic variables {p1, p2, …, pm} A collection of n disjunctive clauses, e.g., ci = (p1 + p2’ + p5) Each clause ci has a corresponding nonnegative weight wi Question What assignment of truth values maximizes the sum of the weights of the satisfied of clauses? See further: 8/23/2019 Copyright G.A. Tagliarini, PhD
11
General Problem Solving Concerns
Goal Representation Constraints (How do I know I have a solution?) “No free lunch theorem” (Wolpert and Macready, 1997) 8/23/2019 Copyright G.A. Tagliarini, PhD
12
Copyright G.A. Tagliarini, PhD
Goal Search Find any solution; time to verify a solution bounded by some polynomial E.g., queen placement, knight’s tour, map coloring, SAT Leads to a Boolean assessment function Optimize Find solutions of a given “quality” (Maximize/Minimize) E.g., TSP, weighted MAX-SAT, WTA, QAP, network flow 8/23/2019 Copyright G.A. Tagliarini, PhD
13
Copyright G.A. Tagliarini, PhD
Representation Choosing a feature space Telling time: rectangular versus polar coordinates Speech recognition: time domain versus frequency domain Face recognition: templates, eigenfaces 8/23/2019 Copyright G.A. Tagliarini, PhD
14
Copyright G.A. Tagliarini, PhD
Constraints How do I know I have a solution? Linear Nonlinear Inequality Integer Zero/One 8/23/2019 Copyright G.A. Tagliarini, PhD
15
Why Can’t I Just Learn One Problem Solving Method?
“No free lunch theorem” (Wolpert and Macready, 1997) No single optimization method will perform best on all instances of all problems. 8/23/2019 Copyright G.A. Tagliarini, PhD
16
Two Conventional Benchmark Techniques
Hill Climbing Simulated Annealing 8/23/2019 Copyright G.A. Tagliarini, PhD
17
Hill Climbing (Minimization)
Assume: Candidate solutions → s0 and s1 in a space S Goal → g Fitness function f such that f:S →R+ Repetition limit → L Algorithm Initialize s0 = s1, f0 = f(s0), f1 = f(s1), Counter While (counter<L, f0>g) s1 = mutate(s0), f1 = f(s1) If f1 < f0 (Note: Improvements are always kept) s0 = s1 f0 = f(s0) Increment counter 8/23/2019 Copyright G.A. Tagliarini, PhD
18
Stochastic Hill Climbing (Minimization)
Assume: Candidate solutions, goal, fitness function, solution space, and iteration limit as in basic hill climbing A scaling parameter T Algorithm Initialize While (counter<L, f0>g) s1 = mutate(s0), f1 = f(s1) Df = (f0 – f1) If rand >1/ (1+exp(Df/T) (Note: Improvements are often kept) s0 = s1 f0 = f(s0) Increment counter 8/23/2019 Copyright G.A. Tagliarini, PhD
19
Copyright G.A. Tagliarini, PhD
Simulated Annealing 8/23/2019 Copyright G.A. Tagliarini, PhD
20
Evolutionary Algorithms
Based on analogy to evolution in biological systems Key process elements Fitness function “Artificial selection” Crossover and recombination Mutation 8/23/2019 Copyright G.A. Tagliarini, PhD
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.