ABSTRACT
This thesis presents the development of smell agent optimization (SAO) algorithm. The developed algorithm consists of three modes (sniffing, trailing and random modes). The evaporation of smell molecules from the smell source is modelled into sniffing mode using the concept of the hydrostatic pressure of gas and positions of molecules. The fitness of the sniffing mode is evaluated and the molecule with the most favourable fitness is taken as the agent. The olfaction capacity of the agent is then evaluated and the training mode is developed using the current position of the agent and the position of the molecules with the current worst fitness. In practical scenarios, it is usually difficult for the agent to account for all the evaporating smell molecules due to the Brownian nature of the smell molecules. This is largely responsible for the agent to getting trapped in a "state of confusion" and consequently leading to the loss of smell trail. To account for this situation in the SAO, a random mode which allows the agent to take a random step in the search space is modelled. The agent evaluates the fitness of the random mode and decides whether to continue its trailing process or to start the entire process of the SAO all over again. This process continues until the object (optimum result) generating the smell is identified. The performance of the developed SAO was evaluated using a total of thirty-nine (39) optimization benchmark functions. Simulations were performed using MATLAB R2017a and results were compared with the results obtained using the fruit fly optimization algorithm (FFOA) and gaseous Brownian motion optimization (GBMO). Results showed that the SAO obtained the best results in twenty-two (22) functions (56.41%) while the FFOA and GBMO obtained the best results in four (4) and seven (7) (10.26% and 17.95%) functions respectively. However, there were similar results in six (6) of the functions (15.38%). The convergence rate of the algorithms was also compared and results showed that the FFOA converged faster than the SAO in all the functions except in one, while the GBMO converged faster than the SAO in 24 of the functions. These convergence results are expected because the computation time in FFOA and GBMO is similar to the computation time required to evaluate one and two modes in SAO respectively. The developed SAO was applied to path planning problem and three scenarios of minimum spanning tree (MST) problem and results were compared with particle swarm optimization (PSO) and smell detection agent (SDA). Though all the algorithms obtained an optimized obstacle free path, results showed that SAO performed better than PSO and SDA in terms of cost by 11.41% and 83.29% respectively. On the MST model, the SAO and PSO obtained the same cost in the first scenario and 3.03% improvement over SDA. In the second scenario, the SAO obtained a better cost with 15.97% and 20.67% improvement over PSO and SDA respectively. In the third scenario, the SAO obtained a better cost with 8.94% and 14.14% improvement over PSO and SDA respectively. These results showed that the developed SAO is highly efficient and can compete significantly well with other algorithms reported in the literature.