Cooperative control and optimization of robotic agents

Abstract

The cooperate behavior that emerges from the interactions among simple multi-agent robots along with the solution possibilities these interactions provide, has formed part of the growing research areas in recent years within the confines of artificial intelligence. In this thesis, we explore these possibilities leveraging single and multi-objective optimization on a machine-learning algorithm: the artificial neural network. The rationale is to achieve goal oriented collaborative control (group behavior) with regard to localizing multiple gradient sources during search operations. With a view to solving the multisource localization problem, we develop an ingenious hybrid metaheuristics algorithm for optimizing exploration (search-oriented) and exploitation (goal-oriented) in both fully observable and partially observable domains. We compared the performance of our model with the existing state of the art metaheuristic algorithms such as Simulated Annealing (SA), Cuckoo-search (CK), Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) using 30 standard benchmark complex objective functions. Results showed significant improvement. The homing (goal-oriented) operation introduced novel concepts for accelerating off-policy reinforcement learning algorithm for Partially Observable Markov Decision Processes (POMDP) via dynamic programming on a multi-agent framework. Finally, we demonstrated an ingenious approach to the resampling phase of Monte Carlo’s particle filter (for robot localization) which showed relatively significant improvement in the belief state estimation accuracy with respect to ground truth within POMDP domains. The contribution of this research is twofold: firstly, it presents a framework for search optimization while localizing multiple emission sources. Secondly, it presents ingenious concepts for foraging, gradient source localization along with the potential for search and rescue operations within POMDP environments.