• Since the 1970s that the idea of a general algorithmic framework, which can be applied with relatively few modifications to different optimization problems, emerged. • Metaheuristics: methods that combine rules and randomness while imitating natural phenomena. • These methods are from now on regularly employed in all the sectors of business, industry, engineering. • besides all of the interest necessary to application of metaheuristics, occasionally a new metaheuristic algorithm is introduced that uses a novel metaphor as guide for solving optimization problems. 2 League Championship Algorithm: A new algorithm for numerical function optimization By: Dr. A. H. Kashan Some examples • particle swarm optimization algorithm (PSO): models the flocking behavior of birds; • harmony search (HS): models the musical process of searching for a perfect state of harmony; • bacterial foraging optimization algorithm (BFOA): models foraging as an optimization process where an animal seeks to maximize energy per unit time spent for foraging; • artificial bee colony (ABC): models the intelligent behavior of honey bee swarms; • central force optimization (CFO): models the motion of masses moving under the influence of gravity; • imperialist competitive algorithm (ICA): models the imperialistic competition between countries; • fire fly algorithm (FA): performs based on the idealization of the flashing characteristics of fireflies. 3 League Championship Algorithm: A new algorithm for numerical function optimization By: Dr. A. H. Kashan Any attempt to design algorithms work on one or several or distributed problem-solving Are inspired by nature’s neighborhood structure(s) devices inspired by the collective capability to evolve living imposed on the members beings well adapted to of the search space.behavior of social insect colonies their environment Metaheuristicsand other animal societies Evolutionary algorithms Trajectory methods Swarm intelligence Social, political, music, sport , etc Tabu search Evolution strategies Ant colony optimization Genetic programming Variable neighborhood Particle swarm optimization search Genetic algorithm Artificial bee colony Bacterial foraging optimization Group search optimizer 4 League Championship Algorithm: A new algorithm for numerical function optimization Harmony search Society and civilization Imperialist competitive algorithm League championship algorithm By: Dr. A. H. Kashan A sports league is an organization that exists to provide a regulated competition for a number of teams to compete in a specific sport. Formations are a method of positioning players on the pitch to allow a team to play according to its pre-set tactics. The main aim of match analysis is: to identify strengths (S) which can then be further built upon, to identify weaknesses (W) which suggest areas for improvement, to use data to try to counter opposing strengths (threats (T)) and exploit weaknesses (opportunities (O)) This kind of analysis is typically known as strengths/weaknesses/opportunities/ threats (SWOT) analysis The SWOT analysis, explicitly links internal (S/W) and external factors (O/T). Identification of SWOTs is essential because subsequent steps in the process of planning for achievement of the selected objective may be derived from the SWOTs. 6 League Championship Algorithm: A new algorithm for numerical function optimization By: Dr. A. H. Kashan In strategic planning there are four basic categories of matches for which strategic alternatives can be considered: S/T matches show the strengths in light of major threats from competitors. The team should use its strengths to avoid or defuse threats. S/O matches show the strengths and opportunities. Essentially, the team should attempt to use its strengths to exploit opportunities. W/T matches show the weaknesses against existing threats. Essentially, the team must attempt to minimize its weaknesses and avoid threats. These strategy alternatives are generally defensive. W/O matches illustrate the weaknesses coupled with major opportunities. The team should try to overcome its weaknesses by taking advantage of opportunities. The SWOT analysis provides a structured approach to conduct the gap analysis. A gap is “the space between where we are and where we want to be”. A transfer is the action taken whenever a player moves between clubs. 7 League Championship Algorithm: A new algorithm for numerical function optimization By: Dr. A. H. Kashan LCA, is a population based algorithmic framework for global optimization over a continuous search space. A common feature among all population based algorithms is that they attempt to move a population of possible solutions to promising areas of the search space, in terms of the problem’s objective, during seeking the optimum. 8 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan Sporting terminology Evolutionary (LCA) terminology League Population week iteration Team i ith member in the population formation solution playing strength fitness value Maximum iterations Number of seasons 9 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan 1) It is more likely that a team with better playing strength wins the game. 2) The outcome of a game is not predictable given known the teams’ playing strength perfectly. It is not unlikely that the world leading FC BARCELONA loses the game to ZORRAT-KARANE-PARS-ABAD from Iranian 3rd soccer division. 3) The probability that team i beats team j is assumed equal from both teams point of view. 4) The outcome of the game is only win or loss (We will later break this rule). 5) Any strength helped team i to win from team j has a dual weakness caused j to lose. In other words, any weakness is a lack of a particular strength. 6) Teams only focus on their upcoming match without regards of the other future matches. Formation settings are done just based on the previous week events. 10 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan f ( X ( x1 , x2 ,..., xn )) : an n dimensional numerical function that should be minimized over the decision space defined by xdmin xd xdmax, d 1,..,n X it ( xit1 , xit2 ,..., xint ) : A formation (a potential solution) for team i at week t f ( X it ) : indicates the fitness/function value resultant from X it Bit (bit1 , bit2 ,..., bint ) : the best formation for team i experienced till week t To determine Bit , a greedy selection is done at each iteration as follows: If f ( X it ) f ( Bit 1 ) Bit X it ; Else if Bit Bit 1 ; End if 11 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan Week 1 Week 2 . . Week L-1 Team 1 Team 2 Team L A League schedule is generated Teams play in pairs based on the league schedule at week t, and winner/ loser are determined using a playing strength based criterion; Terminate YES 1. Through an artificial match analysis, changes 1. t=1 are done in the team 2. initialize team formation (new solution) formations 2. The playing strength 3. initialize best along with the resultant formations formation is determined (fitness calculation) 3. current best formation is updated. Start is t< S×(L-1) ? NO NO Is it the end of the season? YES Week 1 Week 2 Week L-1 Team 1 Team 2 Team L Do possible transfers for each team 13 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan In an ideal league environment we can assume a linear relationship between the team’s playing strength and the outcome of its game. proportional to its playing strength, each team may have a chance to win (idealized rule 2) we determine the winner/loser in a stochastic manner by allowing teams to have their chance of win based on their degree of fit The degree of fit is proportional to the team’s playing strength and is measured based on the distance with an ideal reference point. 14 We assume that a better team can comply with more factors that an ideal team owns. t p Consider teams i and j to fight at week t. Define i as the expected t { f ( Bit )} chance of team i to beat team j at week t and f imin 1,...,L idealized rule 1 f (X ) f t t f (X ) f pi t i t j t p tj pit f ( X tj ) f t f ( X tj ) f ( X it ) 2 f t p p 1 idealized rule 3 Since teams are evaluated based on their distance with a common reference, the ratio of distances determines the winning portions. A random number in [0,1] is generated, if it is less than or equal to pit team i wins and team j losses; otherwise j wins and i losses (idealized rule 4). t i t j 15 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan l= Index of the team that will play with team i based on the league schedule at week t+1. j= Index of the team that has played with team i based on the league schedule at week t. k= Index of the team that has played with team l based on the league schedule at week t. 16 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan No Could we WIN the game from team j at week t ? Yes the loss is directly due to our WEAKNESSES the success is directly due to our STRENGTHES Idealized rule 5 Idealized rule 5 the loss is directly due to the STRENGTHES of team j the success is directly due to the WEAKNESSES of team j Artificial match analysis doing by team i (S/W evaluation) 17 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan Could our opponent WIN the game from team k at week t? No the opponent’s style of play might be a direct OPPORTUNITY Threats are the results of their playing STRENGTHES Idealized rule 5 Focusing on the WEAKNESSES of team k, gives us a way of avoiding the possible threats Idealized rule 5 Focusing on the STRENGTHES of team k, gives us a way of affording the possible opportunities the opponent’s style of play might be a direct THREAT Opportunities are the results of their playing WEAKNESSES Artificial match analysis doing by team i (O/T evaluation) 18 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan S/T strategy S W S/O strategy W/T strategy W/O strategy i was winner i was winner i was loser i was loser l was winner l was loser l was winner l was loser Focusing on … Focusing on … Focusing on … Focusing on … own strengths own strengths (or weaknesses of j) (or weaknesses of j) - - - - own weaknesses own weaknesses (or strengths of j) (or strengths of j) weaknesses of l O (or strengths of k) strengths of l T weaknesses of l - strengths of l - (or weaknesses of k) (or strengths of k) (or weaknesses of k) 19 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan Assume that team k has won the game from team l. To beat l, it is reasonable that team i devises a playing style rather similar to that was adopted by team k at week t . By “ X kt X it ” we address the gap between the playing style of team i and team k, sensed via “focusing on the strengths of team k”. In a similar way we can interpret “ X it X kt ” when “focusing on the weaknesses of team k”. In other words, it may be reasonable to avoid a playing style rather similar to that was adopted by team k. We can interpret “ X tj X it ” or “ X it X tj ” in a similar manner. 20 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan If i was winner and l was winner, then t 1 t t t t t t x b y ( r ( x x ) r ( x x d 1,...,n (S/T equation): id id id 1 1 id kd 1 2 id jd )) Else if i was winner and l was loser, then (S/O equation): xidt 1 bidt yidt ( 2 r1 ( xkdt xidt ) 1r2 ( xidt x tjd )) d 1,...,n Else if i was loser and l was winner, then (W/T equation): xidt 1 bidt yidt ( 1r2 ( xidt xkdt ) 2 r1 ( xtjd xidt )) d 1,...,n Else if i was loser and l was loser, then the (W/O equation): xidt 1 bidt yidt ( 2 r2 ( xkdt xidt ) 2 r1 ( xtjd xidt )) d 1,...,n End if 21 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan In above formulas we rely upon the fact that normally teams play based on their current best formation (that found it suitable over the times), while preparing the required changes recommended by the match analysis. 1 and 2 are constant coefficients used to scale the contribution of “retreat” or “approach” components, respectively. the diversification is controlled by allowing to “retreat” from a solution and also by coefficient 1 , while the intensification is implicitly controlled by getting “approach” to a solution and by coefficient 2 . We refer the above system of updating equations as LCA/recent since they use the teams’ most recent formation as a basis to determine the new formations. 22 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan If i was winner and l was winner, then (S/T equation): xidt 1 bidt yidt ( 1r1 (bidt bkdt ) 1r2 (bidt btjd )) d 1,...,n Else if i was winner and l was loser, then (S/O equation): xidt 1 bidt yidt ( 2 r1 (bkdt bidt ) 1r2 (bidt btjd )) d 1,...,n Else if i was loser and l was winner, then (W/T equation): xidt 1 bidt yidt ( 1r2 (bidt bkdt ) 2 r1 (btjd bidt )) d 1,...,n Else if i was loser and l was loser, then the (W/O equation): xidt 1 bidt yidt ( 2 r2 (bkdt bidt ) 2 r1 (b tjd bidt )) d 1,...,n End if 23 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan It is unusual that coaches do changes in all or many aspects of the team. normally a few number of changes are devised. To simulate the number of changes (q use a truncated geometric distribution. t i n d 1 y id ) made in Bit , we n q0 1 ln( 1 ( 1 ( 1 p ) )r ) t t c qi q 1 : q {q0 , q0 1,..., n} 0 i ln(1 pc ) Where r is a random number in [0,1] and pc (0,1) is a control parameter. q0 is the least number of changes realized during the artificial match analysis t q i number of dimensions are selected randomly from Bit and their value is changed according to one of the Equations 24 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan f1 ( x ) f 2 (x) f 3 ( x) n i 1 x xi [100 ,100 ] 2 i n 1 100 ( xi2 xi 1 ) 2 (1 xi ) 2 i 1 n i 1 ( xi2 10 cos(2xi ) 10 ) f 4 ( x ) 20 exp 0.2 1 . n x i 1 exp( 1 . n n i 1 f 5 ( x ) 418 .9829 n 25 n 2 i xi [2.048 ,2.048 ] xi [5.12,5.12] xi [32.76 , 32.76 ] cos(2xi )) 20 e n i 1 xi sin( xi ) xi [512 .03,511 .97 ] Comparison is done between LCA and the highly recognized (PSO) algorithm L 10 N particles 10 S 1000 N iterations 9000 1 0.5 w 0.9 linear 0.1 2 0.5 c1 2 pC 0.01 c2 2 vmax/ min xmax/ min 26 League Championship Algorithm: A new algorithm for numerical function optimization By: A. H. Kashan 27 Mean of best values for 3 10 4 Mean of best values for 3 Schwefel function 10 Rosenbrock function LCA PSO 2 10 10 LCA PSO 2 10 1 10 1 10 0 0 10 f(X) f(X) 10 -1 10 -1 10 -2 10 -2 10 -3 10 -3 10 -4 10 -4 10 -5 10 -5 10 -6 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10 10000 0 1000 2000 3000 Week/Iteration Rastrigin function 6000 7000 8000 9000 10000 LCA PSO 10 Sphere function LCA PSO Ackley function 1 10 0 Mean of best values for 4 10 1 LCA PSO 2 10 0 10 10 -1 -1 10 0 10 10 -2 f(X) 10 -3 10 -4 10 f(X) -2 f(X) Mean of best values for 2 10 10 5000 Week/Iteration Mean of best values for 2 4000 -3 10 -4 10 -4 10 -5 -2 10 10 -5 10 10 -6 10 -6 -6 10 10 -7 10 28 -7 0 1000 2000 3000 4000 5000 6000 Week/Iteration 7000 8000 9000 10000 10 -8 0 500 1000 1500 Week/Iteration 2000 2500 10 0 500 1000 1500 Week/Iteration 2000 2500 29 Week 1 Week 5 Week 10 Week 20 Week 50 30 Week 100 In order to see that whether each of S/T, S/O, W/T and W/O updating equations has a significant effect on the performance of LCA, we sequentially omit the possible effect that each equation might have on the evolution of the solutions. 31 0 4 10 10 LCA/best/omitting S/T equation LCA/best/omitting S/O equation LCA/best/omitting W/T equation LCA/best/omitting W/O equation LCA/best -1 10 -2 10 LCA/best/omitting S/T equation LCA/best/omitting S/O equation LCA/best/omitting W/T equation LCA/best/omitting W/O equation LCA/best 2 10 0 10 -2 -3 10 -4 10 -5 10 -6 10 -7 10 10 -4 10 -6 10 -8 10 -10 10 -12 10 -8 10 -14 0 200 400 600 800 1000 1200 1400 1600 10 1800 0 50 100 150 Weeks 200 250 300 350 Weeks 4 5 10 10 LCA/best/omitting S/T equation LCA/best/omitting S/O equation LCA/best/omitting W/T equation LCA/best/omitting W/O equation LCA/best 0 10 LCA/best/omitting S/T equation LCA/best/omitting S/O equation LCA/best/omitting W/T equation LCA/best/omitting W/O equation LCA/best 2 10 8 10 0 10 6 -2 10 10 -4 10 -5 10 4 10 -6 10 -8 10 2 10 LCA/best/omitting S/T equation LCA/best/omitting S/O equation LCA/best/omitting W/T equation LCA/best/omitting W/O equation LCA/best -10 10 -10 10 -12 10 -15 10 -14 0 1000 2000 3000 4000 Weeks 32 5000 6000 7000 8000 9000 10 0 1000 2000 3000 4000 0 10 5000 Weeks 6000 7000 8000 9000 0 1000 2000 3000 4000 Weeks 5000 6000 7000 8000 9000 Learning from team’s previous game only If i was winner, then (S equation): xidt 1 bidt yidt ( 1r1 (bidt btjd )) Else if i was loser, then (W equation):xidt 1 bidt yidt ( 2 r1 (btjd bidt )) End if Learning from opponent’s previous game only If l was winner, then t 1 t t t t x b y ( r ( b b (T equation): id id id 1 1 id kd )) Else if l was loser, then t bidt )) (O equation):xidt 1 bidt yidt ( 2 r1 (bkd End if 33 d 1,...,n d 1,...,n d 1,...,n d 1,...,n 4 0 10 10 LCA/best/Learning from team's previous game only LCA/best/Learning from opponent's previous game only LCA/best 2 10 -2 10 0 10 -4 10 -2 10 -6 10 -4 10 -6 10 -8 10 -8 10 -10 10 LCA/best/Learning from team's previous game only LCA/best/Learning from opponent's previous game only LCA/best -12 10 -14 10 0 200 400 600 800 -10 10 -12 10 -14 1000 1200 1400 1600 10 1800 0 50 100 150 200 Weeks 250 300 350 400 450 500 Weeks 4 5 10 10 LCA/best/Learning from team's previous game only LCA/best/Learning from opponent's previous game only LCA/best LCA/best/Learning from team's previous game only LCA/best/Learning from opponent's previous game only LCA/best 2 10 8 10 0 10 0 10 -2 6 10 10 -4 f(X) 10 -5 10 4 10 -6 10 -8 10 2 10 -10 10 LCA/best/Learning from team's previous game only LCA/best/Learning from opponent's previous game only LCA/best -10 10 -12 10 -15 10 -14 0 1000 2000 3000 4000 Weeks 34 5000 6000 7000 8000 10 9000 0 1000 2000 3000 4000 Weeks 0 10 5000 6000 7000 8000 9000 0 1000 2000 3000 4000 Weeks 5000 6000 7000 8000 9000 • Interestingly, these empirical results are in accordance with the business reality. • In business strategy there are two schools of thought, the “environmental (external)” and the “resource based (internal)”. • Through 1970s and 80s, the dominant school was the environmental school which dictates that a firm should analyze the forces present within the environment in order to asses the profit potential of the industry. • Nevertheless, above average performance is more likely to be the result of core capabilities inherent in a firm’s resources (internal view) than its competitive positioning in its industry (external view). 35 Tie outcome is interpreted as the consequent of the strengths/ opportunities and weaknesses/threats 36 Tie outcome is neutral. There is no learning from ties 37 Tie outcome is randomly interpreted as win or loss For example, in this situation, under the case of “Else if i was winner and l had tied” the new formation is set up as follows: xidt 1 bidt yidt ( 1r1ui (bidt bkdt ) 2 r2 (1 ui )(bkdt bidt ) 1r3 (bidt btjd )) Tie outcome is interpreted as win If i had won/tied and l had won/tied, then use (S/T) equation to setup a new formation Else if i had won/tied and l was loser, then use (S/O) equation setup a new formation Else if i was loser and l had won/tied, then use (W/T) equation to setup a new formation Else if i was loser and l was loser, then use (W/O) equation to setup a new formation End if 38 Tie outcome is interpreted as loss If i was winner and l was winner, then use (S/T) equation to setup a new formation Else if i was winner and l had lost/tied, then use (S/O) equation setup a new formation Else if i had lost/tied and l was winner, then use (W/T) equation to setup a new formation Else if i had lost/tied and l had lost/tied, then use (W/O) equation to setup a new formation End if 39 4 0 10 10 LCA/best/win-loss-tie 1 LCA/best/win-loss-tie 2 LCA/best/win-loss-tie 3 LCA/best/win-loss-tie 4 LCA/best/win-loss-tie 5 LCA/best -2 10 -4 10 LCA/best/win-loss-tie 1 LCA/best/win-loss-tie 2 LCA/best/win-loss-tie 3 LCA/best/win-loss-tie 4 LCA/best/win-loss-tie 5 LCA/best 2 10 0 10 -2 10 -4 10 -6 10 -6 10 -8 -8 10 10 -10 10 -10 10 -12 10 -12 10 -14 0 200 400 600 800 1000 1200 1400 1600 10 1800 0 50 100 150 Weeks 200 250 Weeks 4 10 4 LCA/best/win-loss-tie 1 LCA/best/win-loss-tie 2 LCA/best/win-loss-tie 3 LCA/best/win-loss-tie 4 LCA/best/win-loss-tie 5 LCA/best 10 2 10 0 10 LCA/best/win-loss-tie 1 LCA/best/win-loss-tie 2 LCA/best/win-loss-tie 3 LCA/best/win-loss-tie 4 LCA/best/win-loss-tie 5 LCA/best 2 10 8 10 0 10 -2 6 10 10 -2 10 -4 10 -4 10 4 10 -6 10 -6 LCA/best/win-loss-tie 1 LCA/best/win-loss-tie 2 LCA/best/win-loss-tie 3 LCA/best/win-loss-tie 4 LCA/best/win-loss-tie 5 LCA/best 10 -8 10 -8 10 -10 10 -10 10 -12 10 -12 2 10 0 10 10 -14 0 1000 2000 3000 4000 Weeks 40 5000 6000 7000 10 0 1000 2000 3000 4000 Weeks 5000 6000 7000 8000 9000 0 1000 2000 3000 4000 Weeks 5000 6000 7000 8000 900 41 0 5 10 10 LCA/best/Tr=0.1 LCA/best/Tr=0.3 LCA/best/Tr=0.5 LCA/best/Tr=0.7 LCA/best/Tr=0.9 LCA/best -2 10 0 10 -4 10 -6 10 -5 10 -8 10 LCA/best/Tr=0.1 LCA/best/Tr=0.3 LCA/best/Tr=0.5 LCA/best/Tr=0.7 LCA/best/Tr=0.9 LCA/best -10 10 -12 10 -14 10 0 200 400 600 800 1000 Weeks 4 1200 1400 1600 -10 10 -15 1800 10 0 20 40 60 80 4 10 120 140 160 180 200 10 10 2 100 Weeks 10 LCA/best/Tr=0.1 LCA/best/Tr=0.3 LCA/best/Tr=0.5 LCA/best/Tr=0.7 LCA/best/Tr=0.9 LCA/best 2 10 10 8 0 10 -2 -2 10 10 -4 6 10 -4 10 10 -6 4 10 -6 10 10 -8 -8 10 LCA/best/Tr=0.1 LCA/best/Tr=0.3 LCA/best/Tr=0.5 LCA/best/Tr=0.7 LCA/best/Tr=0.9 LCA/best -10 10 -12 10 -14 10 10 0 10 0 42 1000 2000 3000 4000 5000 Weeks 6000 7000 8000 9000 10 LCA/best/Tr=0.1 LCA/best/Tr=0.3 LCA/best/Tr=0.5 LCA/best/Tr=0.9 LCA/best/Tr=0.9 LCA/best -10 10 -12 10 -14 10 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 2 10 0 10 0 1000 2000 3000 4000 5000 Weeks 6000 7000 8000 9000