New answers tagged


it can be either. If you consider the lack of reward as "penalty" then getting 0 reward is bad. if you use a value estimator through a neural network, the range of rewards will dictate the squashing function you use for the output layer

Top 50 recent answers are included