Statistical Analysis of Autonomous Agents' Decisions in Learning to Cross Cellular Automaton Based Highway
We conduct statistical analysis of the performance of a population of autonomous agents learning to cross a cellular automaton based highway. The performance of autonomous agents is measured by various statistics, which are used to analyze agents’ four types of decisions, e.g. (1) correct crossing decisions; (2) incorrect crossing decisions; (3) correct waiting decisions; (4) incorrect waiting decisions. We study how agents’ performance depends on a decision-making formula type the agents use and on the presence of risk takers and of risk avoiders in their population. We investigate the effects of accumulation of prior information in agents’ knowledge base (i.e., of transfer of the knowledge base built in one traffic environment to the agents in another traffic environment) on agents’ success in learning to cross the highway in the population of agents without and with risk takers and risk avoiders. We describe briefly the simulation model and the agents’ two learning algorithms based on a type of “observational social learning” strategy. In this strategy, each agent attempting to cross the highway reviews the outcomes of decisions of the preceding agents, mimics what worked for them and avoids what did not. We develop formal probabilistic description of the simulation model, in terms of various probability spaces and the associated random elements, needed for carrying out our statistical analysis. We describe the simulation experiments setups, the types of collected simulation data and how these data were organized in proper data bases needed for statistical analysis. We conduct statistical analysis of the simulation data by means of entropy, mutual information of random elements, normalized distance between random elements. We investigate means and standard deviations of various statistics of the simulation model. Additionally, we apply several statistical techniques, including regression tree analysis and statistical learning analysis. The obtained statistical results, not only compare the agents’ performance for the two decision-making formulas, but also provide a better understanding of how the key configuration parameters’ values affect the agents’ ability to learn to cross successfully the highway. Based on the presented statistical analysis the improvements to the simulation model are recommended and outlined in this Thesis.