Research On Collective Decision-making In Uncertain Environments

Collective decision-making is one of the fundamental problems in the area of swarm intelligence.The study of collective decision-making problems is of great theoretical significance for understanding the group behavior of biological groups in nature.It also has significant application value in the area of unmanned swarm systems.In collective decision-making problems,the main challenge arises from uncertainty.According to the source of uncertainty,this paper categorizes uncertainty into state uncertainty caused by imperfect perception and model uncertainty caused by imperfect knowledge.State uncertainty means that the agent cannot determine its state and environmental conditions according to the observations,while model uncertainty means that the agent lacks enough knowledge to model the problem and accurately calculates the expected reward of each possible behavior.In order to explain the influence of the interactions among the agents on the decision quality of agents in uncertain environments,this paper first considers collective decisionmaking problems under state uncertainty and analyzes the mechanism and methods of reducing state uncertainty with inter-agent interactions.Collective decision-making problems under model uncertainty are also studied,and the mechanism of agents adapting to new environments faster with the interactions among the agents is investigated.The main contributions of this paper are summarized as follows:First,for the collective decision-making problems under state uncertainty,this paper theoretically analyzes the relationship between the Bayesian collective decision-making model and the Ising model.The similarity between the agents trying to maximize their benefits and the ferromagnetic particles is revealed.Inspired by this result,the generalized Ising model with dynamic confidence(GIM-C)is proposed.Based on the theoretical tools of statistical physics,the reason why GIM-C biases the agents to choose the same behavior is analyzed,and the numerical experiment also verifies this result.Subsequently,by modeling the collective decision-making problem in a known dynamic environment as a partially observable stochastic game problem,this paper theoretically analyzes the effectiveness of the dynamic confidence weighting mechanism.Based on this result,this paper utilizes an evolutionary algorithm to develop a knowledge and data co-driven method for collective decision-making problems.Numerical simulations show that the resulting method achieves high performance in dynamic environments.This result demonstrates that the dynamic confidence weighting mechanism helps the agents in the group to reduce state uncertainty and make better decisions in dynamic environments.Finally,in order to investigate the mechanism of inter-agent interactions benefiting the group in collective decision-making problems under model uncertainty,two collective Thompson sampling algorithms are proposed to solve the Bernoulli collective bandit problem.Based on the theoretical analysis and numerical simulations of the proposed two collective Thompson sampling algorithms,this paper further considers a more general case,i.e.,the sub-Gaussian collective bandit problems,and proposes a new method named the confidence-weighted Boltzmann-Gumbel exploration policy(C-CBGE).Both theoretical analysis and experimental results show that the performance of C-CBGE will improve with the increase in the number of agent neighbors.In summary,this paper develops various decision-making models and methods for collective decision-making problems under state uncertainty and model uncertainty.The mechanism of inter-agent interactions benefiting the group in collective decision-making problems under uncertainty is also revealed by theoretical analysis and numerical simulations.