This paper proposes new algorithms to improve Reinforcement Learning (RL) and Deep Q Network (DQN) methods for path planning considering uncertainty in the environment's perception. The authors aim to formulate and solve the path planning optimization problem by minimizing the path length, avoiding obstacles, and minimizing the related uncertainty. In this regard, a reward function is constructed based on a weighted feature of the environment images, robot and environment constraints, and path optimality criterion as the optimization problem's objective function. In this study, Deep Learning (DL) is used for two purposes. First, for perceiving a real environment to find the state transition matrix of the mobile robot path planning problem, and second, for extracting the state’s features directly from an image of the environment to select the appropriate actions. To solve the path planning problem, it is formed in the context of an RL problem, and a Convolutional Neural Network (CNN) is used to approximate Q-values as a linear parameterized function. Implementing this approach improves the Q-learning, SARSA, and DQN algorithms as the new versions, called POQL, POSARSA, and PODQN. The learning process results show that using newly improved algorithms increases path planning performance by more than 20%, 21%, and 5% compared to the Q learning, SARSA, and DQN, respectively.
This paper proposes an innovative Federated Reinforcement Learning (FRL) approach called the Adaptive Confidence-Weighted Policy Aggregation method, or ACWPA in short. In light of incomplete information and heterogeneous knowledge, ACWPA was developed to combine strengths from multiple agents while canceling their weaknesses in multi-agent tasks. This method dynamically weights the contribution of agents’ policies to provide a global policy based on the agent’s performance and the relevance of their expertise when the information about state-action rewards is partially incomplete. Evaluated on a multi-agent path planning task, ACWPA demonstrates advanced convergence and generalization compared to standard FRL methods like FedAvg and FedProx. Outcomes show that ACWPA increases navigation efficiency by 20% and reduces collision charges by 35% throughout diverse environments, highlighting its capacity to boost collaborative knowledge in multi-agent systems with heterogeneous knowledge. Furthermore, implementing ACWPA on large language models (LLMs) yielded a 15% improvement, indicating that this method has potential applicability in different areas of artificial intelligence.