Heuristic Scheduling Based on Policy Learning - Google Sites
Recommend Documents
production systems is done by allocating priorities to jobs waiting at various machines through these dispatching heuris
machine centres, loading/unloading station and work-in-process storage racks. Five types of parts were processed in the
Diamantino Falcão. GECAD Research Group-School of. Engineering. Polytechnic Institute of Porto. Porto, Portugal [email protected].
Feb 1, 2010 - agent based approach to patient scheduling using EBL is better. ... nature of hospitals because of the proactive and reactive nature of agents.
matematika means Mathematics, informatika means Informatics, e.t.c.. • the string
Mokytojai[i][3] defines the code of the classroom of the i-th teacher, ..... Visa
programin˙e ↩iranga laisvai prieinama tinklapiuose naudojant Java kalb ↩a.
Daug.
Sep 20, 2011 - A second case is when J is only known as a computable function, e.g. when learning a Tetris policy [28]. In such inria-00625001, version 1 - 20 ...
Department of Industrial Engineering and Management Sciences. Northwestern University ... First, for each unit of Work-In-Process Inventory that is waiting at a ...
Technology Management) Thomas Morton Free ... in Engineering and Technology Management) Popular , Heuristic Scheduling S
Dec 19, 2014 - âCentER and Department of Econometrics and Operations Research, ... that produce advanced products comp
paper, we propose a hybrid heuristic-based Artificial Immune System (AIS) algorithm for solving the scheduling problem. In this regard, AIS with some heuristics ...
[5] H. Baghban and M. Rahmani, âA Heuristic on Job Scheduling in Grid Computing Environmentâ, in Proc. 7th Inter. Conf. on Grid and. Cooperative Computing ...
SCHEDULING OF JOB SHOPS USING TABU SEARCH ... Keywords: Scheduling, Multiple Objective Tabu search, Optimization, Dispatching Rules. 1.
Nov 22, 2011 - 1In NVIDIA architectures, 32 threads are executed together in lock-step as a warp and in AMD GPU architec
Jul 20, 2007 - of one compute resource for executing the job and a data host each to access ... note that even in the second case, the data host is considered.
certain fraction of service requests for a hosted application is required to have a specific response time [3]. As geographi- cally distributed data centers, each ...
Bruce A. Draper ...... 11 Lee D. Erman, Frederick Hayes-Roth, Victor R. Lesser, and D. Raj Reddy. The ... 14 Frederick S. Hillier and Gerald J. Lieberman.
In IEEE Real-Time Sys- tems Symposium, pages 112â122, 1985. [4] Carl Edward Rasmusen and Chris Williams. Gaus- sian processes for machine learning.
Today, Vehicular Ad-Hoc Networks (VANETs) offer a promising way to achieve ... where vehicles have non-safety download requests to be served by a battery-powered ... Section V lays out a detailed presentation of the MDP formulation. The.
May 28, 2004 - Computers & Operations Research 32 (2005) 2731â2750 ... Therefore, this research proposes two heuristic approaches that lead to good.
in allocating resources for a large streaming application. 1 Introduction ..... Service. QoS. Monitoring. Service. Application. Information. Service. Grid. Boundary.
Reinforcement learning algorithms [11], [3] are very suit- able for learning to .... AC makes a TD-update to the Critic's value-function V : V (st) := V (st) + β(rt + γV ...
Aug 28, 2009 - Relayed transmission techniques have the advantages of enhancing the ..... concepts for wireless and mobi
a Facultad de IngenierÃa Industrial y Turismo, Universidad Central âMarta Abreuâ de ... investigadores debido a su gran complejidad y por ende la necesidad de ...
Heuristic Scheduling Based on Policy Learning - Google Sites
Dec 5, 2001 - Some Heuristic Rules(Dispatching Rule/ Scheduling. Rules). These are Simple priority rules based on inform
06-06991 Research Skills
Heuristic Scheduling Based on Policy Learning Project Supervisor: Dr. Thorsten Schnier By: Vineet Khare (MSc. Natural Computation)
School of Computer Science University of Birmingham 5 December, 2001 1
Overview •
Introduction 1. Problem 2. Some Heuristic Rules. 3. Why learning?
•
Some Techniques (with Merits & Demerits) 1. Without Knowledge Modification. 2. With Knowledge Modification.
•
Conclusion
2
Introduction- 1 Problem •
Scheduling problem
•
Solution in One Step- Global Optimal Solution
•
Solution using Local Heuristics- Good (enough) Solution “Which batch of components should I process next on this machine?”
•
Various AI techniques 1. Expert Systems 2. Other Methods- Random, Blind, Heuristic. 3
Introduction- 2 Some Heuristic Rules(Dispatching Rule/ Scheduling Rules) These are Simple priority rules based on information available related to jobs. Information Available Processing Time Due dates Slack Arrival Time
Rule Shortest processing time (SPT) Earliest Due Date (EDD) Minimum Slack(MINISLK) First-in first-out (FIFO) etc.
4
Introduction- 3 Why Learning? •The system should be able to adapt to the changing environment. •They must include learning components to correct their misconceptions and improve their performance based on experience. •Difficulty of obtaining experts •Most environments are so dynamic that knowledge becomes obsolete too fast.
5
Some Techniques- 1 •
In General 4 Steps: 1. Obtaining knowledge from sources(Human Expert, Simulation data) 2. Store this knowledge in digital form(e.g. decision trees). 3. Making Decision 4. Modification of stored knowledge.
•
Some Evaluation Criterion 1. Robustness to unforeseen events 2. Performance on benchmark problems 3. Degree to which they are understood by humans.
6
Some Techniques- 2
By By: :H.Fisher, H.Fisher, G.L. Thompson G.L. Thompson 1961 1961
1. Probabilistic Learning Combinations of Local JSS Rules •Which rule to choose (SIO or LRT), at a particular decision point, is selected using a learning process. •Computational experimentation to produce learning how to vary the frequency of use of each of these rules.
+ Easy to implement, better than any one of the 2 rules used -Doesn’t give very good results -Not understandable by humans. 7
Some Techniques- 3
By: By: Quinlan Quinlan1986 1986
2. Iterative Dichotomister 3 (ID 3) •Uses examples to induce production rules (e.g. IF ... THEN ...), which form a simple decision tree.
•Classification: Nodes in a decision tree correspond to attributes of the objects to be classified, and the arcs are alternative values for these attributes. •The end nodes of the tree (leaves) indicate classes to which groups of objects belong. 8
Some Techniques- 4
By By:S.Piramuthu :S.Piramuthu etetal. al.1994 1994
3. Pattern Directed Scheduling
•Decision Support System that adaptively utilizes the most appropriate heuristics given the current state of the system. •Learning by experimentation- Simulation and Inductive Learning modules. •Constructive feedback between these modules provides- knowledge refinement
9
Some Techniques- 5
By By: :S.Piramuthu S.Piramuthu etetal. al.1994 1994
3. Pattern Directed Scheduling (contd.)
+ Knowledge Modification + Computational time required is small enough for using it as a real time aid. 10
Some Techniques- 6
By By: :C.Chiu, C.Chiu, Y.Yih 1994 Y.Yih 1994
4. Learning Based Methodology for Dynamic Scheduling • Includes three modules: 1. Discrete-event simulation, 2. Instance generation: Searching of good training examples
using genetic algorithm.
3. Incremental induction: A tolerance-based incremental learning algorithm to allow continuous learning and facilitate knowledge modification.
+Modifies the knowledge contained in the rule bases. +Explores routing flexibility and handles uncertainties. 11
Some Techniques- 7
By By: : W.Zhang W.Zhang1996 1996
5. Reinforcement Learning Approach •A repair-based scheduler starts with a critical-path schedule and repairs constraint violations. •Temporal difference algorithm TD(λ) is applied to train a neural network to learn a heuristic evaluation function over states. •Reinforcement learning methods learn a policy + Can be used to repair a schedule when some unforeseen difficulty arises. - Formulation of scheduling task is tough. 12
Some Techniques- 8
By By: : Y.Yih Y.Yihetetal. al.1998 1998
6. Integration of Inductive Learning & NN •Developing multi-objective FMS scheduler. •Different classes for different objectives. •Competitive NN associate a simulation output with a class.
•Inductive learning is then used to obtain scheduling rules from this classified data. 13
Some Techniques- 9
By By: : Y.Yih Y.Yihetetal. al.1998 1998
6. Integration of Inductive Learning & NN (contd.)
+ Real time scheduling. + Knowledge modification. + Multi objectives can be handled. - Knowledge acquired can not be explicitly represented.
14
Conclusion
Knowledge modification is important when the scheduling environment is dynamic. Different techniques are well sited with different scheduling scenarios. •e.g if there are multiple objectives – Integration of IL and NN works well. •For handling unforeseen events- Reinforcement learning should be used.