Remember what the current view hides.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Begin with two visible positions

A dot is at 0.35 scene units, then 0.43 one frame later. Its observed change is 0.08 units per frame. If this speed continues, two hidden frames later it should be at 0.59. That calculation uses history; a single blank frame contains neither position nor direction.

The experiment supplies the first two frames and the last frame of an eight-frame episode. Five intervening frames contain a missing flag and a zero coordinate placeholder. Zero does not mean the object was measured at the origin. Training uses eight complete authored episodes; six different starts provide repeated public development. The supervised target is the next position from the synthetic training environment. Development targets never enter updates.

A state that can learn to retain information

The network has three hidden units. At each frame it combines the observed coordinate, visibility flag and its previous three numbers, then applies tanh. A linear readout predicts the next position. All 22 parameters learn together. The hidden state starts at zero for every episode; it is not a saved memory shared across episodes.

Write the observation vector as $o_t$, the three retained numbers as $h_t$, and the prediction as $\hat{x}_{t+1}$. The recurrence is $h_t=\tanh(W_o o_t+W_h h_{t-1}+b)$. The output is $\hat{x}_{t+1}=w^T h_t+c$. Each symbol denotes weights or numbers computed inside the experiment, not hidden simulator truth. Tanh bounds each hidden unit between −1 and +1.

Training minimizes mean half-squared position error. Backpropagation accumulates contributions through all eight frames before the simultaneous update. A direct CPU action makes one or five updates, at most 1,000 total; there is no background timer. The exact hidden values and input flags remain inspectable.

Compare information, not just scores

The current-frame rule repeats an observed position and uses zero when missing. The recent-history rule estimates velocity from the last two visible points and extrapolates. It is strong when constant velocity is correct. A true-state diagnostic has privileged simulator information and zero error by construction; it is not an available deployed baseline.

The recurrent prediction is a point, not a calibrated probability. A separate finite-hypothesis calculation enumerates positive/negative speed and, optionally, a hidden reversal. It retains hypotheses consistent with the visible prefix and reports their minimum and maximum next position. This range is exact only within the declared family.

A missing event cannot be remembered

Turn on a reversal after frame four. The visible prefix is unchanged, so the neural prediction remains unchanged during occlusion. If hidden turns are excluded from the hypothesis family, even its exact range can be wrong. Adding that possibility widens the range: uncertainty reflects missing information, not merely lack of training.

Capture a matched comparison and explain which assumption made each method succeed or fail. Fixed axes disclose clipped predictions; the table retains exact values. These authored episodes and temporary notes do not establish real tracking performance or independent proficiency.