Inferring Reward Functions from State Transitions: A Deep Learning Approach

In the field of reinforcement learning, reward functions are essential for training agents aiming to execute desired actions. However, designing reward functions manually is difficult, time-consuming, and can lead to inaccuracies. This research introduces a novel framework called the Scale In-variant Reward Framework (SIRF) for creating dense rewards using a neural network (NN) classifier. SIRF employs neural networks to mix complex state functions and generate dense bounded reward signals. SIRF's effectiveness is evaluated using the DDPG (Deep Deterministic Policy Gradient) algorithm in an OpenAI Gym environment, a popular open-source toolkit for reinforcement learning research. The results demonstrate that SIRF achieves comparable performance to existing reward functions from the OpenAI Gym environment in terms of key metrics such as average reward, network loss, velocity, and energy consumption for continuous environment agents. This suggests that SIRF can effectively guide the agent's learning process while maintaining consistency in the reward scale, simplifying the overall reinforcement learning setup.