Prominent theories in cognitive science propose that humans understand and represent the knowledge of the world through causal relationships. In making sense of the world, we build causal models in our mind to encode cause-effect relations of events and use these to explain why new events happen by referring to counterfactuals — things that did not happen. In this paper, we use causal models to derive causal explanations of the behaviour of model-free reinforcement learning agents. We present an approach that learns a structural causal model during reinforcement learning and encodes causal relationships between variables of interest. This model is then used to generate explanations of behaviour based on counterfactual analysis of the causal model. We computationally evaluate the model in 6 domains and measure performance and task prediction accuracy. We report on a study with 120 participants who observe agents playing a real-time strategy game (Starcraft II) and then receive explanations of the agents' behaviour. We investigate: 1) participants' understanding gained by explanations through task prediction; 2) explanation satisfaction and 3) trust. Our results show that causal model explanations perform better on these measures compared to two other baseline explanation models.

Prevalent theories in cognitive science propose that humans understand and represent the knowledge of the world through causal relationships. In making sense of the world, we build causal models in our mind to encode cause-effect relations of events and use these to explain why new events happen. In this paper, we use causal models to derive causal explanations of behaviour of reinforcement learning agents. We present an approach that learns a structural causal model during reinforcement learning and encodes causal relationships between variables of interest. This model is then used to generate explanations of behaviour based on counterfactual analysis of the causal model. We report on a study with 120 participants who observe agents playing a real-time strategy game (Starcraft II) and then receive explanations of the agents' behaviour. We investigated: 1) participants' understanding gained by explanations through task prediction; 2) explanation satisfaction and 3) trust. Our results show that causal model explanations perform better on these measures compared to two other baseline explanation models.

Causal explanations present an intuitive way to understand the course of events through causal chains, and are widely accepted in cognitive science as the prominent model humans use for explanation. Importantly, causal models can generate opportunity chains, which take the form of 'A enables B and B causes C'. We ground the notion of opportunity chains in human-agent experimental data, where we present participants with explanations from different models and ask them to provide their own explanations for agent behaviour. Results indicate that humans do in-fact use the concept of opportunity chains frequently for describing artificial agent behaviour. Recently, action influence models have been proposed to provide causal explanations for model-free reinforcement learning (RL). While these models can generate counterfactuals-things that did not happen but could have under different conditions-they lack the ability to generate explanations of opportunity chains. We introduce a distal explanation model that can analyse counterfactuals and opportunity chains using decision trees and causal models. We employ a recurrent neural network to learn opportunity chains and make use of decision trees to improve the accuracy of task prediction and the generated counterfactuals. We computationally evaluate the model in 6 RL benchmarks using different RL algorithms, and show that our model performs better in task prediction. We report on a study with 90 participants who receive explanations of RL agents behaviour in solving three scenarios: 1) Adversarial; 2) Search and rescue; and 3) Human-Agent collaborative scenarios. We investigate the participants' understanding of the agent through task prediction and their subjective satisfaction of the explanations and show that our distal explanation model results in improved outcomes over the three scenarios compared with two baseline explanation models.

Convolutional neural network (CNN) models for computer vision are powerful but lack explainability in their most basic form. This deficiency remains a key challenge when applying CNNs in important domains. Recent work on explanations through feature importance of approximate linear models has moved from input-level features (pixels or segments) to features from mid-layer feature maps in the form of concept activation vectors (CAVs). CAVs contain concept-level information and could be learned via clustering. In this work, we rethink the ACE algorithm of Ghorbani et~al., proposing an alternative invertible concept-based explanation (ICE) framework to overcome its shortcomings. Based on the requirements of fidelity (approximate models to target models) and interpretability (being meaningful to people), we design measurements and evaluate a range of matrix factorization methods with our framework. We find that non-negative concept activation vectors (NCAVs) from non-negative matrix factorization provide superior performance in interpretability and fidelity based on computational and human subject experiments. Our framework provides both local and global concept-level explanations for pre-trained CNN models.

This thesis explores how reinforcement learning (RL) agents can provide explanations for their actions and behaviours. As humans, we build causal models to encode cause-effect relations of events and use these to explain why events happen. Taking inspiration from cognitive psychology and social science literature, I build causal explanation models and explanation dialogue models for RL agents. By mimicking human-like explanation models, these agents can provide explanations that are natural and intuitive to humans.