A Futurist approach to dynamic environments Jano van Hemert, Leiden - PDF document

✬ ✩ A “Futurist” approach to dynamic environments Jano van Hemert, Leiden University, The Netherlands & Clarissa Van Hoyweghen, University of Antwerp, Belgium & Eduard Lukschandl, Ericsson & Hewlett-Packard & Katja Verbeeck, University of Brussels, Belgium presented by Jano van Hemert jvhemert@liacs.nl http://www.liacs.nl/~jvhemert ✫ ✪

✬ ✩ Introduction 2 How it all started Coil Summer School 2000, Limerick, Ireland ☞ People assigned to groups to solve different problems ☞ Conor Ryan provided our group with two dynamic problems ☞ He has attempted to solve those problems using diploid chromosomes ☞ Our objective set was to try to solve them using one of the techniques presented at the summer school ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Introduction 3 How it all started Coil Summer School 2000, Limerick, Ireland ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Introduction 4 The next half hour ① Problem descriptions ② General idea ③ Two tested implementation ④ Experiments & Results ⑤ Conclusions & Future Work ⑥ Questions & Discussion ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Problem descriptions 5 0 − 1 Knapsack – Definition ✔ Goal is to fill a knapsack with objects ✔ Each object has a weight and value assigned ✔ Every 15 generations the maximum allowed weight is changed ✔ Maximum weight is switched between 50% and 80% of the total weight of all the objects ✔ Total of 400 generations (time steps) is used ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

� ✬ ✩ Problem descriptions 6 0 − 1 Knapsack – Behaviour 90 optimal value 80 70 0 50 100 150 200 250 300 350 400 generations ☞ Optimum changes over time ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Problem descriptions 7 O˘ smera’s function — Definition g 1 ( x, t ) = 1 − e 200( x − c ( t )) 2 with c ( t ) = 0 . 04( ⌊ t/ 20 ⌋ ), x ∈ { 0 . 000 , . . . 2 . 000 } , each time step t ∈ { 0 , . . . 1000 } equal to one generation ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

� ✬ ✩ Problem descriptions 8 O˘ smera’s function — Behaviour 0.25 0.2 g(x,t) 1 0.15 c(t) 0.5 0.1 0 0 0.5 0.05 0 1 200 x 400 1.5 600 t 800 0 2 0 20 40 60 80 100 t ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ General idea 9 Predicting the future f(x,t) use predictor for future population 0 t t+ ∆ maxgen regress fitness predictor acquire fitness values ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ General idea 10 Learning from the future migration current future population population ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ General idea 11 Parameters ✔ m determines how many individuals are copied to the current population, best m from the future are selected and overwrite the worst m in the current population ✔ ∆ determines how many generations ahead the future population lives ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Experimental setup 12 Two experiments Perfect prediction ☞ Idea is that the best what could happen is that you have a perfect prediction of the future ☞ With these problems this is very easy to implement as we know exactly the optimum for t + ∆ ☞ If this is not successful, we could ask ourselves if it is useful to continue with the idea of predicting the future Noisy prediction ☞ Could the use of a predictor be harmful? ☞ We give the algorithm noisy and deceptive predictions of the future ☞ Knapsack problem gets wrong optimum (deceptive) and O˘ smera gets a ✫ ✪ random value gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Experimental & Results 13 Experimental setup For both problems we do ✔ a test without any future population ✔ tests with four parameter settings (two pairs) for perfect prediction ✔ tests with four parameter settings (two pairs) for noisy / deceptive predictor ✔ 50 runs for each test with unique random seeds ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Experiments & Results 14 Knapsack results ∆ predictor m error stdev best run none 16.6% 3.52 8.96% × × perfect 5 10 11.9% 3.77 4.85% perfect 15 10 20.3% 4.26 11.7% perfect 5 50 11.8% 3.70 5.97% perfect 15 50 21.4% 6.06 12.0% deceptive 5 10 12.6% 4.12 6.77% deceptive 15 10 13.0% 3.74 7.50% deceptive 5 50 12.7% 4.02 6.47% deceptive 15 50 12.8% 4.07 4.99% ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

� � � ✬ ✩ Experiments & Results 15 Knapsack results 25 "knapsack_clean" 20 15 Error (%) 10 5 0 0 50 100 150 200 250 300 350 400 Generations 25 25 "knapsack_perfect_d5_m10" "knapsack_noisy_d15_m50" 20 20 15 15 Error (%) Error (%) 10 10 5 5 0 0 0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400 ✫ ✪ Generations Generations gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Experiments & Results 16 O˘ smera results ∆ predictor m error stdev best run none 63.8% 10.3 41.2% × × perfect 5 10 0.261% 0.153 0.0751% perfect 5 50 0.168% 0.148 0.0266% perfect 10 10 0.241% 0.220 0.0680% perfect 10 50 0.203% 0.099 0.0698% noisy 5 10 0.241% 0.186 0.0488% noisy 5 50 0.144% 0.122 0.0358% noisy 10 10 0.241% 0.186 0.0488% noisy 10 50 0.168% 0.148 0.0266% ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

� � � ✬ ✩ Experiments & Results 17 O˘ smera results 100 "osmera_clean" 90 80 70 60 Error (%) 50 40 30 20 10 0 0 100 200 300 400 500 600 700 800 900 1000 Generations 0.6 0.6 "osmera_perfect_d5_m50" "osmera_noisy_d10_m50" 0.5 0.5 0.4 0.4 Error (%) Error (%) 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 200 400 600 800 1000 0 200 400 600 800 1000 ✫ ✪ Generations Generations gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Conclusions & Future Research 18 Conclusions Pros and cons ✘ Knapsack problem is better solved with a look-a-head time of 5 generations as opposed to 15, which is the length of the cycle ✔ Adding future predictions when solving the knapsack problem slightly improves the performance when using a deceptive function or when using small values for m ✔ Adding future predictions when solving O˘ smera’s function seems to help ✘ There is little difference in performance between using a perfect or noisy predictor... ✔ There is little difference in performance between using a perfect or noisy predictor... ✫ ✪ gecco -2001 Workshop on Dynamic Optimization

✬ ✩ Conclusions & Future Research 19 Future Research ☞ How sensitive are the parameters m and ∆? ☞ Why does this work well for a real-valued problem and not for a problem from a discrete domain? ☞ Could we replace this whole complicated process by adding more disturbance? For instance with a high mutation rate? ✫ ✪ gecco -2001 Workshop on Dynamic Optimization