Time-series forecasting is widely applied in areas such as energy management, retail demand planning, and operational logistics, where accurate predictions can enhance strategic decision-making. This paper addresses a fundamental challenge in applying transformer architectures to multivariate time-series data: how to preserve inter-feature relationships at each time step when adapting models designed for discrete tokens to continuous numerical sequences. Unlike existing approaches that process each feature dimension independently (column-wise), we propose channel-aware embedding strategies that treat each time step as a unified multivariate observation, preserving the correlations between features within each temporal instance. We systematically evaluate five embedding approaches across two architectural paradigms. First, using GPT-2 as a controlled testbed, we compare three strategies: (1) direct projection of complete time steps into embedding space, (2) windowed tokenization that groups consecutive observations, and (3) adaptive feature binning that creates overlapping sub-vectors to capture local feature correlations while maintaining temporal order. Each approach is tested both with and without Haar wavelet decomposition to separate multi-scale temporal patterns. Second, we develop purpose-built encoder-decoder transformers (DET and ABT) that integrate these channel-aware representations with wavelet processing, demonstrating the transferability of insights from the GPT experiments to more parameter-efficient architectures. On ETTm2 dataset, our bin-based approach reduces MSE by half versus the GPT-based direct embedding for long-horizon forecasts, while our transformer models shows significant improvements on shorter horizons. Our methods also demonstrate strong performance across traffic, weather, and illness datasets.

We address collaborative decision in the Multi-Agent Consistency-based online learning of relational action models. This framework considers a community of agents, each of them learning and rationally acting following their relational action model. It relies on the idea that when agents communicate, on a utility basis, the observed effect of past actions to other agents, this results in speeding up the online learning process of each agent in the community. In the present article, we discuss how collaboration in this framework can be extended to the individual decision level. More precisely, we first discuss how an agent's ability to predict the effect of some action in its current state is enhanced when it takes into account all the action models in the community. Secondly, we consider the situation in which an agent fails to produce a plan using its own action model, and show how it can interact with the other agents in the community in order to select an appropriate action to perform. Such a community aided action selection strategy will help the agent revise its action model and increase its ability to reach its current goal as well as future ones.

Multi Agent Relational Action Learning considers a community of agents, each rationally acting following some relational action model. The observed effect of past actions that led an agent to revise its action model can be communicated, upon request, to another agent, speeding up its own revision. We present a frame-work for such collaborative relational action model revision.

In the Relational Reinforcement learning framework, we propose an algorithm that learns an action model allowing to predict the resulting state of each action in any given situation. The system incrementally learns a set of first order rules: each time an example contradicting the current model (a counter-example) is encountered, the model is revised to preserve coherence and completeness, by using data-driven generalization and specialization mechanisms. The system is proved to converge by storing counter-examples only, and experiments on RRL benchmarks demonstrate its good performance w.r.t state of the art RRL systems.