Isolate the source of DQN training stabilisation

Determine whether the training stabilisation produced by the Spatial Opponent Model under partial observability stems from subgoal inference or from the history-derived structured input.

Background

The paper finds that adding a Spatial Opponent Model to a dueling double DQN stabilises training in partially observable competitive foraging environments, particularly where one-step temporal-difference learning is unstable. A blindfold ablation, which zeros the inferred opponent-goal heatmap at deployment, provides partial evidence that the heatmap carries useful opponent information, but the experiment does not separate the causal contribution of subgoal inference from the broader structured history representation supplied by the module.

Consequently, it remains unresolved whether the observed stabilisation is attributable specifically to inferring the opponent's spatial subgoal or instead to incorporating history-derived structured input into the value network. Resolving this distinction would clarify the mechanism underlying the method's robustness.