Observational Learning with Gated Information: A Lightweight RL Simulation Testbed and Latent Signal Models

Observational learning can be framed as active information sampling: an agent must decide when to spend time watching another agent to reduce uncertainty about when reward will be available. We present a compact simulation testbed for this social-timing problem in which (i) a demonstrator produces brief, stochastic cue bursts, (ii) a short reward window opens after each burst ends, and (iii) the observer can only access social timing features by choosing an explicit Observe action, with imperfect cue detection and a noisy window-remaining estimate. Building on the released multi-algorithm pipeline (dueling double DQN baseline, PPO actor-critic, and a tabular Q-learning control), we add controlled manipulations of social sensing and teacher dynamics: (1) a social-blind observer with the social channel disabled, (2) an impaired tracking observer (reduced detection reliability) with reduced detection reliability and noisier timing estimates, (3) a familiarity variable that gradually improves the observer’s effective sensing with repeated exposure, and (4) teacher palatability conditions that change cue-burst (consumption) duration distributions. We compare learning speed and reward-related latent signals across these conditions by logging behavior and synthesizing candidate internal variables (TD error/RPE, observation-driven value updates, action salience, and information-gain impulses), optionally filtered into signal-like traces and analyzed with peri-event averages. The framework provides a lightweight, code-aligned template for diagnosing observational RL policies and for testing which latent variables best explain event-aligned “teaching signals” under different social-information constraints.