Cooperative Search Techniques

1. Paper review of ENGG*6140 Cooperative Search Techniques Shaw & Vincent ENGG*6140 Optimization for Engineering

2. Outline • Introduction of general Cooperative Search Techniques • Introduction of TECHS approach • Introduction of JSSP (Job-Shop Scheduling problem) • TECHS Approach • Experiments • Related Approach • Conclusion ENGG*6140 Optimization for Engineering

3. Introduction of general Cooperative Search Techniques What is Cooperative Search Technique? • Cooperative search is a parallel strategy for search algorithms where parallelism is obtained by concurrently executing several search programs. • The solution space is implicitly decomposed according to the search strategy of each program. • The programs cooperate by exchanging information on previously explored regions of the solution space. ENGG*6140 Optimization for Engineering

4. Introduction of general Cooperative Search Techniques Why we need Cooperative Search Techniques? • Different search techniques have different strengths and weaknesses. • Genetic Algorithm: good at exploration, but not good for local search • Simulated Annealing: capable of producing near optimal solution, but consumes huge amounts of time. • Tabu search:capable to explore and exploit, but requires lots of parameter tuning ENGG*6140 Optimization for Engineering

5. Introduction of TECHS approach What is TECHS? TECHS is Teams for Cooperative Heterogeneous Search (Multi-agent based approach,Genetic Algorithms & Branch-and-Bound method) ENGG*6140 Optimization for Engineering

6. Introduction of TECHS approach • What are search agents? • A search agent is a search process with additional communication abilities. • Two components in search process: • Search model: consists of a set of possible search states and a transition relation that defines the possible successor states to each state. • Search control: select for a search state exactly one of the possible successor states defined by the transition relation. ENGG*6140 Optimization for Engineering

7. Introduction of TECHS approach What are possible types of information to exchange? • Positive information: the information that helps finding a solution to a given instance of a search problem. • e.g information intended to direct the search towards the search goal. • Negative information: the information that does not lead towards finding a solution, but characterizes attribute values that are definitely not part of a solution. • e.g information intended to avoid possible steps in the search process. ENGG*6140 Optimization for Engineering

8. Introduction of TECHS approach Why we need referees? Since each action taken by an agent during its search results in possible information to exchange, it is very important to limit the information that is really exchanged. Referees: Send-referees & Receive-referees ENGG*6140 Optimization for Engineering

9. Introduction of TECHS approach The usage of Referees: Send-referee: select information of all types on the side of the sending agent. Receive-referee: provide an additional filter on the side of the receiving agent. By adding send- and receive-referees and communication channels between the processes, we get search agents. ENGG*6140 Optimization for Engineering

10. Introduction of TECHS approach Criteria of Send-referees: • Success-driven: • Parts of search states that enable good transitions should be preferred to parts that only enabled bad ones. • Demand-driven: • Rating the information to select w.r.t the control strategy of other agents or w.r.t certain attributes solutions should have or not have. • A further criteria: • very unlike that the other agent will find the information on its own. ENGG*6140 Optimization for Engineering

11. Introduction of TECHS approach About data format: In order to communicate information of all types, data format are needed to represent information. This format can be used for communication by all systems. The search systems internally might use totally different data structure. ENGG*6140 Optimization for Engineering

12. Introduction of TECHS approach Criteria of Receive-referees: • Try to measure the impact of the information on the further search of the agent. • This is especially important for positive information, since it tends to broaden the search space if it is not useful. • Filtering negative information is not so critical, but if an agent accumulates too much negative information then the selection process for the next search steps can become time consuming. ENGG*6140 Optimization for Engineering

13. Introduction of TECHS approach Two phases: • Working phase: the agent concentrates on its search. • Cooperation phase: Send-referees select information to be sent away and Receive-referees filters the information received from the other agents. If all agents use the same time interval for working phases one can organize the information interchange in a synchronized manner, else it has to be asynchronous. ENGG*6140 Optimization for Engineering

14. Introduction of TECHS approach How to use exchanged information? • Directly be included in the search state of an agent. • Or it can influence the control the agent employs in its search. (as parameters of the control) ENGG*6140 Optimization for Engineering

15. Introduction of TECHS approach Summary of building a TECHS-based search team • Data format for the different types of information must be found. • The used search systems must be modified to provide send-referees with information. • The search systems must also be modified to enable the integration of the information selected by their receive-referees into search state and search control. (very hard) • Send- and receive-referees must be developed. ENGG*6140 Optimization for Engineering

16. Introduction of JSSP Objective: To solve JSSP (Job-Shop Scheduling problem) Method: TECHS GA GA GA Branch-and-bound Branch-and-bound GA GA ENGG*6140 Optimization for Engineering

17. Introduction of JSSP • Definition: There are n jobs and m machines; each job comprises a set of operations which must each be done on a different machine for different specified time. • Features: • very important practical problem • being among the worst members of class of NP-complete problems. • It is hard for conventional search-based methods to find near-optima in reasonable time. ENGG*6140 Optimization for Engineering

18. Introduction of JSSP An example ENGG*6140 Optimization for Engineering

19. Introduction of JSSP Existed methods to solve JSSP: • Job-Shop System • NY ENGG*6140 Optimization for Engineering

20. Introduction of JSSP Jop-Shop System: • B&B system for solving the JSS problem, implement in C. • B&B based search uses and-trees to represent search states. Transitions : • close leaves in such a tree, if they represent a solution or have a bound that is not better than the best currently known solution. • expand one leaf of the current tree by adding new leaves. ENGG*6140 Optimization for Engineering

21. Introduction of JSSP NY NY is a GA based on the idea of Nakano and Yamada which employs a special version of the Giffler-Thompson algorithm both to create schedules and in the crossover and mutation operators. ENGG*6140 Optimization for Engineering

22. TECHS Approach Central data format: Partial Schedule Full Schedule & Partial Schedule: • Full schedule: contains for each operation of each job the exact start time and that fulfills all requirements on the problem. • Partial schedule: consists of a set of ordering conditions that are based on two relations between operations on a machine. • op1 > op2: op1 is performed before op2 (maybe other operations between them) • op1 op2: op2 directly follows op1. ENGG*6140 Optimization for Engineering

23. TECHS Approach The Send-Referees • Select information of all types out of the current search state and the search sequence produced so far. • Mainly base their selection on the success the pieces of information produced by the agent. • Receiving agents determine the type of the send-referee. ENGG*6140 Optimization for Engineering

24. TECHS Approach B&BB&B • B&B agents exchange positive and negative information to be integrated into the search state. • Send-referee always selects the best known full schedule as the positive information, if there has been some change since the last cooperation phase. • Negative information are partial schedules describing closed subtrees of the current state. The more nodes this subtree has, the more important it is to communicate this information. ENGG*6140 Optimization for Engineering

25. TECHS Approach B&BNY • Agents using a GA can only use positive information to be included into the search state. • Send-referee selects the partial schedules that have the best bound-values as the positive information(and have not been communicated already). ENGG*6140 Optimization for Engineering

26. TECHS Approach NYB&B • Agents using a GA can only produce positive information. B&B agents can use the information both as the control information and as information to be integrated into the search state. • The positive information to be integrated into the search state is the best individual of the NY agent. • Positive control information are individuals representing very good solutions. In addition, they should represent different areas of the search tree of the receiving agent than its currently focused area(they should very different). ENGG*6140 Optimization for Engineering

27. 130 130 130 145 Solution from GA 125 150 130 150 Positive control information ENGG*6140 Optimization for Engineering

28. TECHS Approach NYNY • The only type of information exchanged are positive individuals to be integrated into the current search state. • Criteria for the selection are quality of the solutions and their difference. • This send-referee was not used in the experiments. ENGG*6140 Optimization for Engineering

29. TECHS Approach The receive-referees • Filter incoming information in order to select information that meet the current needs of their search agents. NY Agent • Only receives partial schedules to be integrated into its search state. • receive-referee extend partial schedules to full ones. • Criteria used for selection are quality and difference. ENGG*6140 Optimization for Engineering

30. TECHS Approach B&B Agent • Receive-referee filter positive and negative information to be integrated and positive control information. • Positive full schedules : only the one with the highest quality that is higher than the quality of the best known solution so far is selected.(How does the agent use this info?) • Negative partial schedules: used by the agent to close a leaf. • How? ENGG*6140 Optimization for Engineering

31. Experiments • Benchmarks stemming from the OR-library.(URL: http://mscmga.ms.ic.ac.uk/info.html) • Each experiment, maximal run-time : 28800 seconds(8 hours) • Working phases:125 seconds. • Communication was synchronized using the function of the UNIX socket concept. • GA agent used 2600 population size. ENGG*6140 Optimization for Engineering

32. Solutionqualitycomparison of the single agents Vs TECHS team 20*10 15*15 20*10 • TECH team produced better schedules than all its agents working alone. The team is always better than its best member. ENGG*6140 Optimization for Engineering

33. Comparison of times needed to find comparable solutions • In most case, ETCHS produce comparable solution in less time. ENGG*6140 Optimization for Engineering

34. Related Approaches • Most parallelization approaches based on some cooperation of search agents employ homogeneous agents. • Injection island GA • TEAMWORK • Only a few approaches employ search agents with different search models(heterogeneous agents). • TECHS ENGG*6140 Optimization for Engineering

35. Conclusion • TECHS presented a cooperation of evolution algorithms with systems on other search paradigms. • TECHS exchange different types of information between agents. • Send/Receive-referees reduces the amount of communication while still selecting the important information. • TECHS results in synergetic improvements both in quality of the solutions found and time needed to find solutions of a certain quality. ENGG*6140 Optimization for Engineering

36. Question? ENGG*6140 Optimization for Engineering

37. Solution from outside: 143 150 Positive full schedule for B&B agent 148 135 140 150 150 145 ENGG*6140 Optimization for Engineering

38. 100 110 148 135 148 140 150 150 150 150 145 150 150 Negative Information for B&B Agent ENGG*6140 Optimization for Engineering