The project is located in the area of sequential decision making in Artificial Intelligence (AI). AI planning traditionally treats decision making as a single-shot process: a user specifies a model, an initial state, and a goal, and the planner returns one plan. This workflow is poorly suited to settings where users have preferences that are difficult to state in advance, want to compare alternatives, or need to understand why certain choices are possible or necessary. If a plan is unsuitable, the usual remedy is to revise the model and restart planning. The objective of this project is to replace this process with an iterative dialogue between the user and the planning system. The project covers two closely connected research lines. Research line 1: Interactive exploration and refinement of plan spaces. Rather than returning a single solution, we will develop methods that expose meaningful choices in the space of possible plans. Users will be able to iteratively enforce or prohibit actions, intermediate properties, or groups of decisions and immediately inspect the consequences. We will build on faceted reasoning, where a facet represents a choice that occurs in some solutions but not in all. Facets provide a compact interface to a potentially very large plan space: after each user decision, the system updates the remaining alternatives and identifies which choices remain open. We will investigate incremental algorithms, compact representations, and preprocessing techniques that make sequences of such queries fast enough for interactive use. We will also study bounded, optimal, and cost-sensitive plan spaces and how user feedback can be retained across solver calls. Research line 2: Model-based attention and explanation. An interactive system must decide which of the many available choices to present. We will develop model-based attention mechanisms that use the planning model and the structure of its solution space to prioritize relevant decisions. Facet significance, robustness under restrictions, causal relevance, and the effect of a choice on the remaining plans will direct attention to consequential actions, rules, and model features. A particular focus will be symbolic action policies learned from planning data. Although such policies are nominally interpretable, their feature expressions and rule interactions can be difficult to understand. By encoding policies and planning models in answer-set programming, we will enable users to inspect which rules justify a transition, identify rules that incorrectly remove valid solutions, and explain policy failures. This extends the preliminary framework in the accompanying paper from automated diagnosis toward a general interactive analysis tool. The methods will be evaluated on public benchmark collections from the International Planning Competition and on publicly available learned policies. We will measure coverage and runtime, response times over sequences of interactive queries, reductions of the plan space, the stability of highlighted choices, and the ability to locate known modelling or policy errors. Controlled single-core runs will ensure reliable algorithm comparisons, while large numbers of independent benchmark runs will be executed concurrently on the requested infrastructure.