Academia.eduAcademia.edu

Figure 6 – uploaded by Sandor Markon

See full PDFdownloadDownload figure
Figure 5.6. Default (left) and optimized NN controller performance on a wide range of arrival rates and server capacities on an S-ring with 20 floors and 6 servers. Each point represents a simulation run with 1 million steps. Lower values are better.

Figure 5 6. Default (left) and optimized NN controller performance on a wide range of arrival rates and server capacities on an S-ring with 20 floors and 6 servers. Each point represents a simulation run with 1 million steps. Lower values are better.

Related Figures (6)

Figure 5.1. The S-ring as an elevator system. The sites are numbered from 0 to n—1. Metaheuristics: Progress as Real Problem Solvers
Figure 5.1. The S-ring as an elevator system. The sites are numbered from 0 to n—1. Metaheuristics: Progress as Real Problem Solvers
Table 5.1. The triple in the first column represents the state of the actual site: Cus- tomer waiting, server present, and server present on the next floor. The probability of a state change to the state in the fourth column is given in the second column Columns three and five denote the decision and the reward respectively. I.e., the server has to make a decision (to ‘take’ or to ‘pass’ the customer) if there is a cus- tomer waiting (1xx), and if there is a server present on the same floor (11x) but nc server on the next floor (110). the unobservable case. Next we will introduce some elementary policies: — The most obvious heuristic policy is the greedy one: When given the choice, always serve the customer: 79(0) = 1. Rather counter-intuitively, this policy is not optimal, except in the heavy traffic (p > 0.5) case. This means that a good policy must bypass some customers occasionally. The greedy-policy does not take any information about the state of the sys- tem into account.
Table 5.1. The triple in the first column represents the state of the actual site: Cus- tomer waiting, server present, and server present on the next floor. The probability of a state change to the state in the fourth column is given in the second column Columns three and five denote the decision and the reward respectively. I.e., the server has to make a decision (to ‘take’ or to ‘pass’ the customer) if there is a cus- tomer waiting (1xx), and if there is a server present on the same floor (11x) but nc server on the next floor (110). the unobservable case. Next we will introduce some elementary policies: — The most obvious heuristic policy is the greedy one: When given the choice, always serve the customer: 79(0) = 1. Rather counter-intuitively, this policy is not optimal, except in the heavy traffic (p > 0.5) case. This means that a good policy must bypass some customers occasionally. The greedy-policy does not take any information about the state of the sys- tem into account.
Figure 5.2. Event graph. The thin arrow represents an event at the beginning (ar- rival) scheduling an event at the end of the arrow (departure) immediately. The heavy arrows indicate that the event at the beginning of the arrow may schedule the event at the end of the arrow.
Figure 5.2. Event graph. The thin arrow represents an event at the beginning (ar- rival) scheduling an event at the end of the arrow (departure) immediately. The heavy arrows indicate that the event at the beginning of the arrow may schedule the event at the end of the arrow.
Figure 5.3. Function depart. The probability of an arrival event is set to 0.5.
Figure 5.3. Function depart. The probability of an arrival event is set to 0.5.
Figure 5.4. Schematic representation of algorithm based validation. A regression model describes the functional relationship between the parameter setting of an al- gorithm and its expected performance. Two optimization problems are equivalent, if there are no interactions between the problem and model parameters.
Figure 5.4. Schematic representation of algorithm based validation. A regression model describes the functional relationship between the parameter setting of an al- gorithm and its expected performance. Two optimization problems are equivalent, if there are no interactions between the problem and model parameters.
Figure 5.5. Two NN controllers dealing with different update methods (sequential, random, quasi-parallel) and arrival rates on an S-ring with 20 floors and 6 servers. Left hand: Default controller, all weights are 1.0, right hand: ES-optimized controller. Each point represents a simulation run with 1 million steps. Lower values are better.
Figure 5.5. Two NN controllers dealing with different update methods (sequential, random, quasi-parallel) and arrival rates on an S-ring with 20 floors and 6 servers. Left hand: Default controller, all weights are 1.0, right hand: ES-optimized controller. Each point represents a simulation run with 1 million steps. Lower values are better.
Related topics:
Markov Decision ProcessEvolutionary AlgorithmDiscrete Event Simulation
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved