This paper describes several new online model-free reinforcement learning (RL) algorithms. The aim is to compare these algorithms experimentally with existing algorithms, namely: Q-learning, Sarsa, Actor-Critic, QV-learning, and ACLA.  We designed 4 new reinforcement algorithms, namely: QV2, QVMAX, QVMAX2, and Sarsa+.We show experiments on five maze problems of varying complexity, the first problem is simple, but the other four maze tasks are of a dynamic or partially observable nature. Furthermore, we show experimental results on the cart pole balancing problem.