Kevin Godin-Dubois

Kevin Godin-Dubois, Anil Yaman, Anna V. Kononova

While Central Pattern Generators (CPGs) and Multi-Layer Perceptrons (MLP) are widely used paradigms in robot control, few systematic studies have been performed on the relative merits of large parameter spaces. In contexts where input and output spaces are small and performance is bounded, having more parameters to optimize may actively hinder the learning process instead of empowering it. To empirically measure this, we submit a given robot morphology, with limited proprioceptive capabilities, to controller optimization under two bio-inspired paradigms (CPGs and MLPs) with evolutionary- and reinforcement- trainer protocols. By varying parameter spaces across multiple reward functions, we observe that shallow MLPs and densely connected CPGs result in better performance when compared to deeper MLPs or Actor-Critic architectures. To account for the relationship between said performance and the number of parameters, we introduce a Parameter Impact metric which demonstrates that the additional parameters required by the reinforcement technique do not translate into better performance, thus favouring evolutionary strategies.

Kevin Godin-Dubois, Karine Miras, Anna V. Kononova

Traditional approaches to training agents have generally involved a single, deterministic environment of minimal complexity to solve various tasks such as robot locomotion or computer vision. However, agents trained in static environments lack generalization capabilities, limiting their potential in broader scenarios. Thus, recent benchmarks frequently rely on multiple environments, for instance, by providing stochastic noise, simple permutations, or altogether different settings. In practice, such collections result mainly from costly human-designed processes or the liberal use of random number generators. In this work, we introduce AMaze, a novel benchmark generator in which embodied agents must navigate a maze by interpreting visual signs of arbitrary complexities and deceptiveness. This generator promotes human interaction through the easy generation of feature-specific mazes and an intuitive understanding of the resulting agents' strategies. As a proof-of-concept, we demonstrate the capabilities of the generator in a simple, fully discrete case with limited deceptiveness. Agents were trained under three different regimes (one-shot, scaffolding, interactive), and the results showed that the latter two cases outperform direct training in terms of generalization capabilities. Indeed, depending on the combination of generalization metric, training regime, and algorithm, the median gain ranged from 50% to 100% and maximal performance was achieved through interactive training, thereby demonstrating the benefits of a controllable human-in-the-loop benchmark generator.

Hüseyin Aydın, Kevin Godin-Dubois, Libio Goncalvez Braz et al.

Reinforcement learning (RL) offers a general approach for modeling and training AI agents, including human-AI interaction scenarios. In this paper, we propose SHARPIE (Shared Human-AI Reinforcement Learning Platform for Interactive Experiments) to address the need for a generic framework to support experiments with RL agents and humans. Its modular design consists of a versatile wrapper for RL environments and algorithm libraries, a participant-facing web interface, logging utilities, deployment on popular cloud and participant recruitment platforms. It empowers researchers to study a wide variety of research questions related to the interaction between humans and RL agents, including those related to interactive reward specification and learning, learning from human feedback, action delegation, preference elicitation, user-modeling, and human-AI teaming. The platform is based on a generic interface for human-RL interactions that aims to standardize the field of study on RL in human contexts.