A Reward Function Using Image Processing for a Deep Reinforcement Learning Approach Applied to the Sonic the Hedgehog Game

Research in the Deep Reinforcement Learning (DRL) field has made great use of video game environments for benchmarking performance over the last few years. Most researches advocate for learning through high-dimensional sensory inputs (i.e. images) as observations in order to simulate scenarios that more closely approach reality. Though, when using these video game environments, the common practice is to provide the agent a reward signal calculated through accessing the environment's internal state. However, this type of resource is hardly available when applying DRL to real-world problems. Thus, we propose a reward function that uses only the images received as observations. The proposal is evaluated in the Sonic the Hedgehog game. We analyzed the agent's learning capabilities of the proposed reward function and, in most cases, its performance is similar to that obtained when accessing the environment's internal state.