In this work, we use optimal control to change the behavior of a deep reinforcement learning policy by optimizing directly in the policy's latent space. We hypothesize that distinct behavioral patterns, termed behavioral modes, can be identified within certain regions of a deep reinforcement learning policy's latent space, meaning that specific actions or strategies are preferred within these regions. We identify these behavioral modes using latent space dimension-reduction with Pairwise Controlled Manifold Approximation Projection (PaCMAP). Using the actions generated by the optimal control procedure, we move the system from one behavioral mode to another. We subsequently utilize these actions as a filter for interpreting the neural network policy. The results show that this approach can impose desired behavioral modes in the policy, demonstrated by showing how a failed episode can be made successful and vice versa using the lunar lander reinforcement learning environment.

Understanding the behavior of deep reinforcement learning (DRL) agents is crucial for improving their performance and reliability. However, the complexity of their policies often makes them challenging to understand. In this paper, we introduce a new approach for investigating the behavior modes of DRL policies, which involves utilizing dimensionality reduction and trajectory clustering in the latent space of neural networks. Specifically, we use Pairwise Controlled Manifold Approximation Projection (PaCMAP) for dimensionality reduction and TRACLUS for trajectory clustering to analyze the latent space of a DRL policy trained on the Mountain Car control task. Our methodology helps identify diverse behavior patterns and suboptimal choices by the policy, thus allowing for targeted improvements. We demonstrate how our approach, combined with domain knowledge, can enhance a policy's performance in specific regions of the state space.

We investigate the effect of including application knowledge about a robotic system states’ causal relations when generating explanations of deep neural network policies. To this end, we compare two methods from explainable artificial intelligence, KernelSHAP, and causal SHAP, on a deep neural network trained using deep reinforcement learning on the task of controlling a lever using a robotic manipulator. A primary disadvantage of KernelSHAP is that its explanations represent only the features’ direct effects on a model’s output, not considering the indirect effects a feature can have on the output by affecting other features. Causal SHAP uses a partial causal ordering to alter KernelSHAP’s sampling procedure to incorporate these indirect effects. This partial causal ordering defines the causal relations between the features, and we specify this using application knowledge about the lever control task. We show that enabling an explanation method to account for indirect effects and incorporating some application knowledge can lead to explanations that better agree with human intuition. This is especially favorable for a real-world robotics task, where there is considerable causality at play, and in addition, the required application knowledge is often handily available.

This paper deals with robotic lever control using Explainable Deep Reinforcement Learning. First, we train a policy by using the Deep Deterministic Policy Gradient algorithm and the Hindsight Experience Replay technique, where the goal is to control a robotic manipulator to manipulate a lever. This enables us both to use continuous states and actions and to learn with sparse rewards. Being able to learn from sparse rewards is especially desirable for Deep Reinforcement Learning because designing a reward function for complex tasks such as this is challenging. We first train in the PyBullet simulator, which accelerates the training procedure, but is not accurate on this task compared to the real-world environment. After completing the training in PyBullet, we further train in the Gazebo simulator, which runs more slowly than PyBullet, but is more accurate on this task. We then transfer the policy to the real-world environment, where it achieves comparable performance to the simulated environments for most episodes. To explain the decisions of the policy we use the SHAP method to create an explanation model based on the episodes done in the real-world environment. This gives us some results that agree with intuition, and some that do not. We also question whether the independence assumption made when approximating the SHAP values influences the accuracy of these values for a system such as this, where there are some correlations between the states.