Struct LearningEngine

Source

pub struct LearningEngine { /* private fields */ }

Expand description

The HELM orchestration hub that bundles all six learning sub-modules.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;

let mut engine = LearningEngine::new(LearningConfig::default());
assert!(engine.q_table().is_empty());
assert_eq!(engine.episode_count(), 0);

Implementations§

Source §

impl LearningEngine

Source

pub fn new(config: LearningConfig) -> Self

Constructs a new LearningEngine from a LearningConfig.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;

let engine = LearningEngine::new(LearningConfig::default());
assert_eq!(engine.episode_count(), 0);

Source

pub fn config(&self) -> &LearningConfig

Returns a reference to the current configuration.

Source

pub fn q_table(&self) -> &QTable

Returns a reference to the Q-table.

Source

pub fn reset_epsilon(&mut self, epsilon: f64)

Resets the Q-table exploration epsilon.

Source

pub fn episode_count(&self) -> usize

Returns the total number of completed episodes.

Source

pub fn total_reward(&self) -> f64

Returns the accumulated total reward.

Source

pub fn record_step( &mut self, signal: &CognitionSignal, state: u64, action: ActionKey, next_state: u64, )

Records one ThinkLoop step into all active learning sub-modules.

This should be called inside ThinkLoop::run after each step produces a CognitionSignal.

§Arguments

signal - The signal produced by this ThinkLoop iteration.
state - FNV-1a state key for the current query.
action - The action applied to produce signal.query.
next_state - FNV-1a state key for the observation after the action.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;
use lmm_agent::cognition::learning::q_table::{ActionKey, QTable};
use lmm_agent::cognition::signal::CognitionSignal;

let mut engine = LearningEngine::new(LearningConfig::default());
let sig = CognitionSignal::new(0, "a b".into(), "a b c".into(), 1.0, 0.0);
let s = QTable::state_key("a b");
let s2 = QTable::state_key("a b c");
engine.record_step(&sig, s, ActionKey::Narrow, s2);
assert!(!engine.q_table().is_empty());

Source

pub fn recommend_action( &mut self, state: u64, goal: &str, step: usize, ) -> ActionKey

Selects the recommended action for state using the Q-table and meta priors.

If a meta-adapter warm-start is available, it is added as a prior bias to the greedy Q selection before applying ε-greedy exploration.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;
use lmm_agent::cognition::learning::q_table::{ActionKey, QTable};

let mut engine = LearningEngine::new(LearningConfig::default());
let s = QTable::state_key("hello world");
let action = engine.recommend_action(s, "hello world", 0);
assert!(ActionKey::all().contains(&action));

Source

pub fn end_of_episode( &mut self, cold: &ColdStore, index: &mut KnowledgeIndex, goal: &str, avg_reward: f64, )

Finalises a completed episode: distils knowledge, stores meta-prototype, synthesises PMI facts, and decays ε.

§Arguments

cold - The agent’s cold store after drain_to_cold.
index - The agent’s knowledge index for distillation output.
goal - The natural-language goal that was pursued.
avg_reward - Mean reward across all steps in this episode.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;
use lmm_agent::cognition::memory::ColdStore;
use lmm_agent::cognition::knowledge::KnowledgeIndex;

let mut engine = LearningEngine::new(LearningConfig::default());
let cold = ColdStore::default();
let mut idx = KnowledgeIndex::new();
engine.end_of_episode(&cold, &mut idx, "test goal", 0.5);
assert_eq!(engine.episode_count(), 1);

Source

pub fn federate(&mut self, snapshot: &AgentSnapshot)

Merges a remote AgentSnapshot into the local Q-table (federated step).

Only active when LearningMode::Federated is enabled.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;
use lmm_agent::cognition::learning::q_table::{ActionKey, QTable};
use lmm_agent::types::AgentSnapshot;

let mut engine = LearningEngine::new(LearningConfig::default());
let remote_qt = QTable::new(0.1, 0.9, 0.0, 1.0, 0.0);
let snap = AgentSnapshot { agent_id: "remote".into(), q_table: remote_qt, total_reward: 1.0 };
engine.federate(&snap);
assert_eq!(engine.aggregator().merge_count, 1);

Source

pub fn export_snapshot(&self, agent_id: impl Into<String>) -> AgentSnapshot

Exports an AgentSnapshot for federated sharing with other agents.

§Examples

use lmm_agent::cognition::learning::engine::LearningEngine;
use lmm_agent::cognition::learning::config::LearningConfig;

let engine = LearningEngine::new(LearningConfig::default());
let snap = engine.export_snapshot("agent-1");
assert_eq!(snap.agent_id, "agent-1");

Source