lex-learning-rate

Adaptive learning rate management for LegionIO agents. Part of the LegionIO cognitive architecture extension ecosystem (LEX).

lex-learning-rate maintains per-domain learning rates that adapt to the agent's prediction performance. Wrong predictions raise the rate (more learning needed), correct predictions lower it (stable enough), and surprise or error events provide temporary boosts. Rates mean-revert toward a default over time. Rolling accuracy windows give a per-domain reliability signal.

Key capabilities:

• Per-domain rates: independent learning rate per knowledge domain
• Prediction-driven adjustment: wrong prediction +0.03, correct prediction -0.02
• Event boosts: surprise +0.05, error +0.04
• Mean reversion: rates drift back toward 0.15 default via decay
• Rate labels: very_slow / slow / moderate / fast / very_fast per rate value

Add to your Gemfile:

gem 'lex-learning-rate'

Or install directly:

gem install lex-learning-rate

require 'legion/extensions/learning_rate'

client = Legion::Extensions::LearningRate::Client.new

# Record prediction outcomes
client.record_prediction(domain: :networking, correct: false)
client.record_prediction(domain: :networking, correct: false)

# Record surprise or error events
client.record_surprise(domain: :networking)

# Check current rate
result = client.current_rate(domain: :networking)
# => { domain: :networking, rate: 0.27, label: :fast, accuracy: 0.3 }

# Find domains with highest/lowest learning rates
client.fastest_domains(limit: 3)
client.slowest_domains(limit: 3)

# Feed tick results (auto-extracts prediction outcomes)
client.update_learning_rate(tick_results: tick_phase_results)

# Stats
client.learning_rate_stats

bundle install
bundle exec rspec
bundle exec rubocop

MIT