Added the n-step capability from "The Effect of Multi-step Methods on Overestimation in Deep Reinforcement Learning" (https://arxiv.org/abs/2006.12692) to hopefully add flexibility to approve performance.
Adds a new argument "horizon" to dictate the n-step look-ahead horizon. This defaults to 1, which retains the current behavior of DDPG and TD3. Increasing horizon to integers > 1 will use the new nStepBuffer. The performance has been tested on Pendulum to ensure that nothing is broken, but any improvement in performance on other problems has not been tested/demonstrated.
Added the n-step capability from "The Effect of Multi-step Methods on Overestimation in Deep Reinforcement Learning" (https://arxiv.org/abs/2006.12692) to hopefully add flexibility to approve performance.
Adds a new argument "horizon" to dictate the n-step look-ahead horizon. This defaults to 1, which retains the current behavior of DDPG and TD3. Increasing horizon to integers > 1 will use the new nStepBuffer. The performance has been tested on Pendulum to ensure that nothing is broken, but any improvement in performance on other problems has not been tested/demonstrated.