E. Dovgan, M. Javorski, T. Tušar, B. Filipič.

**Dealing with comfort as an objective in multiobjective optimization of driving strategies**. In Proceedings of the 15th International Multiconference Information Society - IS 2012, vol. A, pp. 103-106, 2012.

Show Abstract |

PDF
When a person drives a vehicle along a route, he/she optimizes the traveling time and the fuel consumption. The same problem is tackled by the Multiobjective Optimization algorithm for discovering Driving Strategies (MODS) which we designed and implemented. However, the driving strategies found with MODS change the control actions frequently (more frequently than humans) and, therefore, the driving comfort is reduced. To improve the driving comfort, we introduced it as an objective in MODS, thus obtaining the Multiobjective Optimization algorithm for discovering Comfortable Driving Strategies (MOCDS). The two algorithms were compared on data from a real-world route and the results show that MOCDS finds highly comfortable driving strategies, especially when the fuel consumption is reduced. On the other hand, when the traveling time is reduced, MODS already finds comfortable driving strategies that cannot be additionally improved.

comfortable driving strategies, multiobjective optimization, traveling time, fuel consumption, driving discomfort
E. Dovgan, M. Javorski, M. Gams, B. Filipič.

**A two-level approach for discovering driving strategies according to conflicting objectives**. In Proceedings of the 14th International Multiconference Information Society - IS 2011, vol. A, pp. 41-44, 2011.

Show Abstract |

PDF
This paper presents a two-level multiobjective optimization approach for discovering driving strategies. The algorithm at the lower level is a deterministic multiobjective algorithm that searches for driving strategies with respect to two conflicting objectives: traveling time and fuel consumption. The algorithm at the upper level is an evolutionary algorithm that searches for the best input data for the lower-level algorithm. This paper describes the improvements of the lower-level algorithm that we presented in the past. Moreover, it presents the upper-level algorithm and the preliminary tests of the two-level approach. Finally, ideas for future work are given.

driving strategies, multiobjective optimization, traveling time, fuel consumption

E. Dovgan.

**Multiobjective genetic discovery of driving strategies**. In Proceedings of the 2nd Jožef Stefan International Postgraduate School Students' Conference - IPSSC 2010, pp. 24, 2010.

Show Abstract |

PDF
Vehicle driving consumes time and energy (fuel, electricity etc.). Usually both have to be minimized. Minimizing the consumption of one of them leads to increasing the consumption of the other. To find driving strategies that take into consideration both objectives, we have implemented a multiobjective genetic algorithm that constructs driving strategies as sets of rules. Optimal sets of rules consist of nondominated solutions and therefore cannot be sorted based on quality since each solution represents a particular trade-off between the two objectives. The final strategy selection is done by the user who uses higher-level information to select the most preferred strategy from the found best solutions.

To test the strategies, we implemented a vehicle simulator. It is defined with the engine, transmission, aerodynamics, braking and wheel characteristics. It simulates driving on a predefined route that consists of segments. Each segment is defined with its length, inclination, radius and velocity limit. Vehicle driving is controlled with a strategy consisting of a set of rules. Each rule has the following form: IF vehicle characteristics INSIDE interval1 AND segment characteristics INSIDE interval2 THEN USE throttle percentage AND gear OR braking percentage.

The implemented multiobjective genetic algorithm is based on NSGA-II. It has the characteristics of genetic algorithms as follows. It randomly initializes a set of driving strategies. Then these strategies are improved over generations where in each generation pairs of strategies are randomly selected, their information is exchanged, their rules are randomly changed, a randomly selected rule is removed and a randomly created rule is added, and the strategies are finally evaluated and added to the set of strategies. This is done in such a way that each strategy is selected once on average in each generation. In addition to these classical genetic algorithm mechanisms, the NSGA-II has dedicated mechanisms in order to meet the multiobjective algorithm requirements: in addition to minimization of the objectives, it preserves the diversity of the strategies with respect to the objectives. This is done using the non-dominated sorting and the crowding distance mechanisms known from the NSGA-II.

Using the described algorithm, we performed preliminary numerical experiments in multiobjective discovery of driving strategies for several predefined simple routes. The results in the form of nondominated sets of solutions are promising and stimulative for further investigation in this problem domain. Future work will include improvements of the algorithm efficiency, its testing on the routes defined with real data, and comparison of our results with the results of other algorithms.

driving strategies, multiobjective optimization, traveling time, fuel consumption
Award:

**Best Poster Award**