spires_train_online

Perform a single online learning step, updating the readout weights toward a target vector using the current reservoir state.

Signature

spires_status spires_train_online(spires_reservoir *r,
                                  const double *target_vec,
                                  double lr);

Parameters

Returns

spires_status — SPIRES_OK on success. On failure:

• SPIRES_ERR_INVALID_ARG — r or target_vec is NULL, or lr is non-positive.

Example

double lr = 1e-3;
 
for (size_t t = 0; t < series_length; t++) {
    /* Step the reservoir */
    spires_step(r, &input[t * num_inputs]);
 
    /* Update weights toward the target for this timestep */
    spires_status s = spires_train_online(r, &target[t * num_outputs], lr);
    if (s != SPIRES_OK) {
        fprintf(stderr, "online training failed at t=%zu\n", t);
        break;
    }
}

Notes

• Update rule. This function implements a delta rule (Widrow-Hoff) update: W_out += lr * (target - y) * x^T, where x is the current reservoir state and y = W_out * x is the current output. This is a gradient descent step on the squared error.
• No allocation. This function performs no heap allocation. It operates entirely on the reservoir’s pre-allocated internal buffers.
• Comparison with ridge. Online training is useful for adaptive or non-stationary tasks where the target distribution changes over time. For fixed datasets, spires_train_ridge typically produces better results because it solves the linear regression problem in closed form.
• Weight initialization. Before the first online update, W_out is zero-initialized (as set by spires_reservoir_create). If you have previously called spires_train_ridge, those weights are used as the starting point for online fine-tuning.
• Thread safety. Do not call this function on a reservoir that is concurrently in use by another thread.

See Also

• spires_train_ridge — batch training via ridge regression.
• spires_step — advance the reservoir before each online update.
• spires_compute_output — compute output without updating weights.
• Online Delta Rule — theoretical background.