nevo.core.optimiser

NEVO Optimiser

Main optimiser class integrating all neuromorphic components.

trs2o(v, lb, ub)

Transform from v-space [-1, 1] to original space [lb, ub].

Parameters:

• v (np.ndarray) – Solution in v-space

• lb (np.ndarray) – Lower bounds in original space

• ub (np.ndarray) – Upper bounds in original space

Returns:

x – Solution in original space

Return type:

np.ndarray

class NEVOptimiser(objective_function, bounds, dimension, population_size=50, memory_size=25, operators=None, operator_mode='trad', neurons_per_ensemble=100, dt=0.001, epsilon=0.1, learning_rate=0.4, td_gamma=0.99, td_lambda=0.0, td_enabled=True, td_learning_rule=None, td_value_model=None, seed=None)

Bases: object

Neuromorphic Evolutionary Optimiser.

Uses basal ganglia circuits to adaptively select between multiple optimisation operators based on the current search state.

__init__(objective_function, bounds, dimension, population_size=50, memory_size=25, operators=None, operator_mode='trad', neurons_per_ensemble=100, dt=0.001, epsilon=0.1, learning_rate=0.4, td_gamma=0.99, td_lambda=0.0, td_enabled=True, td_learning_rule=None, td_value_model=None, seed=None)

Initialize NEVO optimiser.

Parameters:

• objective_function (Callable) – Function to minimize: f(x) -> float Input x is in original space [lb, ub]

• bounds (tuple) – (lower_bounds, upper_bounds) as numpy arrays or scalars

• dimension (int) – Problem dimension

• population_size (int) – Number of candidates per operator per timestep

• memory_size (int) – Size of solution memory (archive)

• operators (List[Operator], optional) – Custom operators (uses default if None)

• operator_mode (str) – Operator selection mode: “trad” | “traditional”, “nm_dual”, or “nm_softmix”

• neurons_per_ensemble (int) – Neurons per ensemble in neural networks

• dt (float) – Simulation timestep (seconds)

• epsilon (float) – Epsilon-greedy exploration rate

• learning_rate (float) – Learning rate for utility weight adaptation

• td_gamma (float) – Discount factor for TD learning

• td_lambda (float) – Lambda parameter for TD(lambda)

• td_enabled (bool) – Enable TD-based value learning in selector

• td_learning_rule (LearningRule, optional) – Custom TD learning rule instance

• td_value_model (ValueModel, optional) – Custom TD value model instance

• seed (int, optional) – Random seed for reproducibility

evaluate(v)

Evaluate objective function from v-space.

Parameters:

v (np.ndarray) – Solution in v-space [-1, 1]

Returns:

fitness – Objective function value

Return type:

float

update_memory(candidates, fitness)

Update memory with new candidates (competitive update).

Parameters:

• candidates (np.ndarray) – Candidate solutions (population_size, dimension)

• fitness (np.ndarray) – Fitness values (population_size,)

build_model()

Build Nengo model.

run(time, verbose=True, use_dl=False)

Run optimisation for specified time.

Parameters:

• time (float) – Simulation time (seconds)

• verbose (bool) – Print progress information

• use_dl (bool) – Use NengoDL backend for GPU acceleration (requires nengo-dl and tensorflow)

print_results()

Print optimisation results.

get_best_solution()

Get the best solution found.

Return type:

tuple

Returns:

• x_best (np.ndarray) – Best solution in original space

• f_best (float) – Best fitness value

get_statistics()

Get optimisation statistics.

Returns:

stats – Dictionary containing various statistics

Return type:

Dict[str, Any]

reset_td_episode()

Start a fresh TD episode without rebuilding the network.

set_td_lambda(lambda_coeff)

Pass-through TD(lambda) configuration.

set_td_learning_rule(learning_rule)

Pass-through learning-rule swap for basal ganglia TD.

set_td_value_model(value_model)

Pass-through value-model swap for basal ganglia TD.