lc_population (faber_neuroevolution v1.2.4)

LC Population Manager - Evolves Liquid Conglomerate controller TWEANNs.

This gen_server manages a population of LC controller neural networks that compete to be the active controller. Each controller is evaluated over a "trial period" (e.g., 5000 evaluations) and rewarded based on:

1. Fitness improvement velocity (faster learning = better) 2. Training efficiency (fewer evaluations to reach target) 3. Avoiding premature convergence (not getting stuck at low fitness) 4. Resource efficiency (low memory/CPU usage)

Evolution Strategy

- Population of N LC TWEANNs (default: 10) - Each generation: trial period for each controller - Active controller is the champion from last generation - After each trial, compute reward using lc_reward:compute_trial_reward/1 - Select top 50% as parents, mutate to create offspring - Repeat

Time Constants

- Trial period: 5000 evaluations (tau_l0 * 5) - Generation: N * trial_period evaluations - Champion persistence: 3 generations (give good controllers time)

Summary

Functions

get_active_controller(Pid)

Get the active LC controller's recommendations. This is called by task_silo to get hyperparameters from the LC TWEANN.

get_recommendations(Pid, SensorInputs)

Get hyperparameter recommendations from the active LC TWEANN. Sensors are fed to the network, outputs are scaled to hyperparameter ranges.

get_state(Pid)

Get current population state for monitoring.

handle_call(Request, From, State)

handle_cast(Msg, State)

handle_info(Info, State)

init(Config)

report_metrics(Pid, Metrics)

Report evolution metrics for reward computation. Called by task_silo after each cohort/generation.

reset(Pid)

Reset population (start fresh evolution).

start_link()

Start LC population manager with default config.

start_link(Config)

Start LC population manager with custom config.

terminate(Reason, State)