An Agent-Based Modeling Simulation of Mobile Ad Hoc Networks

This dissertation presents two novel approaches to modeling and analyzing Mobile Ad Hoc Networks (MANETs) as a human-machine system. The proposed methods are: a fuzzy ω - calculus model and (b) a fuzzy decentralized Markov decision process (FMDP). With the FMDP, a "selfish competitive" algorithm in which each agent exhibits a rational behavior is developed for a context-based game. Here, each agent chooses to play the game in any context that gives the maximum payoff and uses the reward to achieve its intended goal. The results show that, (a) Agents can exhibit dominant behaviors either within a cluster or across context clusters as they seek to satisfy their goals; (b) The context cluster and rewards structures of agents within a cluster have effects on the agents overall expected reward and goal achievement.;On extending the FMDP to a MANET with multiple nodes, it is observed that agents do compete to gain a maximum reward in action selections. It was noted that agents that are willing to cooperate and share information will dominate action selection with most reward. An example in this dissertation are Agents 1, 5, and 6 who show cooperation in selection and execution of a particular action, while Agents 2 and 3 cooperate to execute another action.;Fuzzy ω- calculus was used to obtain MANET sociometric data. The effects of agent actions such as location change are captured during the simulation. The results indicate that (a) There are significant differences between mean trust scores among agents, (b) There are significant differences between mean information sharing scores among agents, (c) There are no significant differences among agents in how the cooperate, trust, and situation awareness and (d) The overall sociometric scores within the experimental MANET show some interaction.;Sensitivity analysis was conducted on the MANET sociometric algorithms using learning and adaptation models. Some relationships between learning and adaptation were observed. For example Agent 10 had the worst decrease in perception of other agents on the sociometric score during learning and it had the worst follower in adapting to the environmental changes and interaction with other agents. We noticed that the "agent-follow-agent (AFA)" algorithm for adaptation is sensitive to individual agent learning since agents who do not follow other agents through information sharing and situation awareness are less likely to trust and cooperate with other agents. Specifically, sociometric scores for agent learning are not independent.