VBAF.Enterprise.NLInterface.ps1
|
#Requires -Version 5.1 <# .SYNOPSIS Phase 13 - Natural Language Interface .DESCRIPTION Trains a DQN agent to interpret and route natural language commands to the correct VBAF enterprise subsystem. The agent observes intent signals extracted from command text and learns when to: - Respond : answer with information, no system action needed (action 0) - Execute : run a single agent action directly (action 1) - Orchestrate: coordinate multiple agents for the request (action 2) - Escalate: ambiguous or high-risk — request human review (action 3) .NOTES Part of VBAF - Phase 13 Enterprise Automation Engine Phase 13: Natural Language Interface PS 5.1 compatible Real data: $Host.UI, Read-Host, command intent classification #> # ============================================================ # PHASE 13 - NATURAL LANGUAGE INTERFACE # ============================================================ class NLInterfaceEnvironment { # State: 4 normalised intent signal dimensions (0.0 - 1.0) # state[0] = SeverityNorm — direct action signal (proven pattern) [double] $SeverityNorm # CurrentSeverity/3.0 [double] $IntentConfidence # 0=ambiguous 1=crystal clear intent [double] $RiskLevel # 0=read-only query 1=destructive action [double] $Complexity # 0=single step 1=multi-agent pipeline [int] $CorrectActions [int] $MissedEscalations [int] $Steps [double] $TotalReward [int] $EpisodeCount # Confusion matrix [int] $TruePositives [int] $FalsePositives [int] $TrueNegatives [int] $FalseNegatives [int] $CurrentSeverity # raw 0-3 (maps directly to optimal action) # Required by VBAF framework [int] $StateSize = 4 [int] $ActionSize = 4 # Step() stores result here — avoids PSCustomObject type corruption (PS 5.1) [double] $LastReward = 0.0 [bool] $LastDone = $false NLInterfaceEnvironment() { $this.Reset() | Out-Null } # CRITICAL PS 5.1: build strictly typed [double[]] element by element [double[]] GetState() { [double[]] $s = @(0.0, 0.0, 0.0, 0.0) $s[0] = $this.SeverityNorm $s[1] = $this.IntentConfidence $s[2] = $this.RiskLevel $s[3] = $this.Complexity return $s } [double[]] Reset() { $this.Steps = 0 $this.TotalReward = 0.0 $this.CorrectActions = 0 $this.MissedEscalations = 0 $this.TruePositives = 0 $this.FalsePositives = 0 $this.TrueNegatives = 0 $this.FalseNegatives = 0 $this.LastDone = $false # CRITICAL: must reset here $this.EpisodeCount++ $this._SampleCondition() [double[]] $initState = $this.GetState() return $initState } [void] _SampleCondition() { # Balanced training distribution # 25% query (0), 30% execute (1), 25% orchestrate (2), 20% escalate (3) $roll = Get-Random -Minimum 1 -Maximum 100 if ($roll -le 25) { $this.CurrentSeverity = 0 } elseif ($roll -le 55) { $this.CurrentSeverity = 1 } elseif ($roll -le 80) { $this.CurrentSeverity = 2 } else { $this.CurrentSeverity = 3 } # SeverityNorm = direct action signal in state[0] [double[]] $snArr = @(0.0) $snArr[0] = $this.CurrentSeverity $snArr[0] /= 3.0 $this.SeverityNorm = $snArr[0] # Generate NL intent metrics consistent with command type switch ($this.CurrentSeverity) { 0 { # Simple query: high confidence, low risk, low complexity $this.IntentConfidence = [double](Get-Random -Minimum 80 -Maximum 100) / 100.0 $this.RiskLevel = [double](Get-Random -Minimum 0 -Maximum 10) / 100.0 $this.Complexity = [double](Get-Random -Minimum 0 -Maximum 15) / 100.0 } 1 { # Execute: clear intent, moderate risk, single agent $this.IntentConfidence = [double](Get-Random -Minimum 60 -Maximum 85) / 100.0 $this.RiskLevel = [double](Get-Random -Minimum 10 -Maximum 40) / 100.0 $this.Complexity = [double](Get-Random -Minimum 15 -Maximum 45) / 100.0 } 2 { # Orchestrate: moderate confidence, higher risk, multi-agent $this.IntentConfidence = [double](Get-Random -Minimum 40 -Maximum 65) / 100.0 $this.RiskLevel = [double](Get-Random -Minimum 40 -Maximum 70) / 100.0 $this.Complexity = [double](Get-Random -Minimum 45 -Maximum 80) / 100.0 } 3 { # Escalate: ambiguous intent, high risk, high complexity $this.IntentConfidence = [double](Get-Random -Minimum 0 -Maximum 40) / 100.0 $this.RiskLevel = [double](Get-Random -Minimum 70 -Maximum 100) / 100.0 $this.Complexity = [double](Get-Random -Minimum 80 -Maximum 100) / 100.0 } } } [int] _OptimalAction() { # 0=Respond 1=Execute 2=Orchestrate 3=Escalate return $this.CurrentSeverity } [void] Step([int]$action) { $this.Steps++ $optimal = $this._OptimalAction() # Symmetric distance-based reward (proven in Phases 10-12) [int] $dist = $action - $optimal if ($dist -lt 0) { $dist = -$dist } if ($dist -eq 0) { $this.LastReward = 2.0; $this.CorrectActions++ } elseif($dist -eq 1) { $this.LastReward = -1.0 } elseif($dist -eq 2) { $this.LastReward = -2.0 } else { $this.LastReward = -3.0 } if ($this.CurrentSeverity -ge 2 -and $action -lt 2) { $this.MissedEscalations++ } $isCritical = ($this.CurrentSeverity -ge 2) $agentActs = ($action -ge 2) if ($isCritical -and $agentActs) { $this.TruePositives++ } if (!$isCritical -and $agentActs) { $this.FalsePositives++ } if (!$isCritical -and !$agentActs) { $this.TrueNegatives++ } if ($isCritical -and !$agentActs) { $this.FalseNegatives++ } $this.TotalReward += $this.LastReward $this._SampleCondition() $this.LastDone = ($this.Steps -ge 200) } } # ------------------------------------ # Real Windows NL command probe # ------------------------------------ function Get-VBAFNLSnapshot { [CmdletBinding()] param() Write-Host "" Write-Host " Probing NL command environment..." -ForegroundColor Gray try { # Show PS version and host info Write-Host (" PowerShell version : {0}" -f $PSVersionTable.PSVersion) -ForegroundColor White Write-Host (" Host name : {0}" -f $Host.Name) -ForegroundColor White # Sample intent classification — keyword-based routing demo $sampleCommands = @( "show me the security events", "restart the network adapter", "optimize all pipelines and alert on failures", "delete all logs and reset everything now" ) $intentLabels = @("Respond", "Execute", "Orchestrate", "Escalate") Write-Host "" Write-Host " Sample command routing:" -ForegroundColor Gray for ($i = 0; $i -lt $sampleCommands.Count; $i++) { Write-Host (" [{0,-12}] {1}" -f $intentLabels[$i], $sampleCommands[$i]) -ForegroundColor DarkCyan } Write-Host "" Write-Host " NL routing capability : confirmed ✅" -ForegroundColor Green } catch { Write-Host " [WARNING] NL probe incomplete: $($_.Exception.Message)" -ForegroundColor Yellow Write-Host " [INFO] Training will use simulated intent signals." -ForegroundColor Gray } } # ============================================================ # MAIN TRAINING FUNCTION # ============================================================ function Invoke-VBAFNLInterfaceTraining { param( [int] $Episodes = 100, [int] $PrintEvery = 10, [switch] $FastMode, [switch] $SimMode, [switch] $SkipRealData ) Write-Host "" Write-Host "💬 VBAF Enterprise - Phase 13: Natural Language Interface" -ForegroundColor Cyan Write-Host " Training DQN agent on NL command routing..." -ForegroundColor Cyan Write-Host " Actions: 0=Respond 1=Execute 2=Orchestrate 3=Escalate" -ForegroundColor Yellow Write-Host " State : SeverityNorm | Confidence | RiskLevel | Complexity" -ForegroundColor Yellow Write-Host " Reward : +2 correct -1 dist=1 -2 dist=2 -3 dist=3" -ForegroundColor Yellow Write-Host "" if (-not $SkipRealData) { Get-VBAFNLSnapshot } $nlEnv = [NLInterfaceEnvironment]::new() # Phase 1: Baseline — inline random loop Write-Host " Phase 1: Baseline (random agent - 10 episodes)..." -ForegroundColor Gray $baseRewards = @() for ($b = 1; $b -le 10; $b++) { $nlEnv.Reset() | Out-Null $bReward = 0.0 while (-not $nlEnv.LastDone) { $rAction = Get-Random -Minimum 0 -Maximum 4 $nlEnv.Step($rAction) $bReward += $nlEnv.LastReward } $baseRewards += $bReward } [double[]] $bAvgArr = @(0.0) $bAvgArr[0] = ($baseRewards | Measure-Object -Average).Average Write-Host (" Baseline avg reward: {0:F2}" -f $bAvgArr[0]) -ForegroundColor Gray if ($FastMode) { $Episodes = [Math]::Min($Episodes, 30) } Write-Host "" Write-Host " Phase 2: Training DQN agent ($Episodes episodes)..." -ForegroundColor Gray # DQN setup $config = [DQNConfig]::new() $config.StateSize = 4 $config.ActionSize = 4 $config.EpsilonDecay = 0.9995 $config.EpsilonMin = 0.05 [int[]] $arch = @(4, 24, 24, 4) $mainNetwork = [NeuralNetwork]::new($arch, $config.LearningRate) $targetNetwork = [NeuralNetwork]::new($arch, $config.LearningRate) $memory = [ExperienceReplay]::new($config.MemorySize) $agent = [DQNAgent]::new($config, $mainNetwork, $targetNetwork, $memory) $results = [System.Collections.Generic.List[object]]::new() for ($ep = 1; $ep -le $Episodes; $ep++) { [double[]] $state = @(0.0, 0.0, 0.0, 0.0) if ($SimMode) { $roll = Get-Random -Minimum 1 -Maximum 100 if ($roll -le 25) { $nlEnv.CurrentSeverity = 0 } elseif ($roll -le 55) { $nlEnv.CurrentSeverity = 1 } elseif ($roll -le 80) { $nlEnv.CurrentSeverity = 2 } else { $nlEnv.CurrentSeverity = 3 } [double[]] $snArr = @(0.0) $snArr[0] = $nlEnv.CurrentSeverity $snArr[0] /= 3.0 $nlEnv.SeverityNorm = $snArr[0] switch ($nlEnv.CurrentSeverity) { 0 { $nlEnv.IntentConfidence = [double](Get-Random -Minimum 80 -Maximum 100) / 100.0 $nlEnv.RiskLevel = [double](Get-Random -Minimum 0 -Maximum 10) / 100.0 $nlEnv.Complexity = [double](Get-Random -Minimum 0 -Maximum 15) / 100.0 } 1 { $nlEnv.IntentConfidence = [double](Get-Random -Minimum 60 -Maximum 85) / 100.0 $nlEnv.RiskLevel = [double](Get-Random -Minimum 10 -Maximum 40) / 100.0 $nlEnv.Complexity = [double](Get-Random -Minimum 15 -Maximum 45) / 100.0 } 2 { $nlEnv.IntentConfidence = [double](Get-Random -Minimum 40 -Maximum 65) / 100.0 $nlEnv.RiskLevel = [double](Get-Random -Minimum 40 -Maximum 70) / 100.0 $nlEnv.Complexity = [double](Get-Random -Minimum 45 -Maximum 80) / 100.0 } 3 { $nlEnv.IntentConfidence = [double](Get-Random -Minimum 0 -Maximum 40) / 100.0 $nlEnv.RiskLevel = [double](Get-Random -Minimum 70 -Maximum 100) / 100.0 $nlEnv.Complexity = [double](Get-Random -Minimum 80 -Maximum 100) / 100.0 } } $nlEnv.CorrectActions = 0 $nlEnv.MissedEscalations = 0 $nlEnv.Steps = 0 $nlEnv.TotalReward = 0.0 $nlEnv.LastDone = $false $nlEnv.EpisodeCount++ $state = $nlEnv.GetState() } else { $state = $nlEnv.Reset() } $done = $false $epReward = 0.0 $respondCount = 0 $executeCount = 0 $orchestrateCount = 0 $escalateCount = 0 [int] $stepCount = 0 while (-not $done) { $action = $agent.Act($state) $nlEnv.Step($action) [double[]] $nextState = $nlEnv.GetState() [double] $reward = $nlEnv.LastReward [bool] $isDone = $nlEnv.LastDone $agent.Remember($state, $action, $reward, $nextState, $isDone) $stepCount++ if ($stepCount % 4 -eq 0) { $agent.Replay() } $state = $nextState $done = $isDone $epReward += $reward switch ($action) { 0 { $respondCount++ } 1 { $executeCount++ } 2 { $orchestrateCount++ } 3 { $escalateCount++ } } } $agent.EndEpisode($epReward) $results.Add(@{ Episode = $ep Reward = $epReward Respond = $respondCount Execute = $executeCount Orchestrate = $orchestrateCount Escalate = $escalateCount Epsilon = $agent.Epsilon }) if ($ep % $PrintEvery -eq 0) { $lastN = $results | Select-Object -Last $PrintEvery $avgSum = 0.0 foreach ($r2 in $lastN) { $avgSum += $r2.Reward } [double[]] $avgArr = @(0.0) $avgArr[0] = $avgSum $avgArr[0] /= $lastN.Count $avg = [Math]::Round($avgArr[0], 2) Write-Host (" Ep {0,4}/{1} AvgReward: {2,7} Eps: {3:F3} Res:{4} Exe:{5} Orc:{6} Esc:{7}" -f ` $ep, $Episodes, $avg, $agent.Epsilon, $respondCount, $executeCount, $orchestrateCount, $escalateCount) -ForegroundColor White } } # Phase 3: Evaluation — inline loop (epsilon=0) Write-Host "" Write-Host " Phase 3: Final evaluation (epsilon=0 - 10 episodes)..." -ForegroundColor Gray $agent.Epsilon = 0.0 $trainedRewards = @() for ($t = 1; $t -le 10; $t++) { [double[]] $evalState = $nlEnv.Reset() $tReward = 0.0 while (-not $nlEnv.LastDone) { $tAction = $agent.Act($evalState) $nlEnv.Step($tAction) [double[]] $evalState = $nlEnv.GetState() $tReward += $nlEnv.LastReward } $trainedRewards += $tReward } [double[]] $tAvgArr = @(0.0) $tAvgArr[0] = ($trainedRewards | Measure-Object -Average).Average Write-Host (" Trained avg reward: {0:F2}" -f $tAvgArr[0]) -ForegroundColor Green [double[]] $impArr = @(0.0) if ($bAvgArr[0] -ne 0) { $impArr[0] = $tAvgArr[0] - $bAvgArr[0] $impArr[0] /= [Math]::Abs($bAvgArr[0]) $impArr[0] *= 100.0 } $bAvg = [Math]::Round($bAvgArr[0], 2) $tAvg = [Math]::Round($tAvgArr[0], 2) $improvement = [Math]::Round($impArr[0], 1) # Precision / Recall [double[]] $precArr = @(0.0) [double[]] $recArr = @(0.0) $denomP = $nlEnv.TruePositives + $nlEnv.FalsePositives $denomR = $nlEnv.TruePositives + $nlEnv.FalseNegatives if ($denomP -gt 0) { $precArr[0] = $nlEnv.TruePositives; $precArr[0] /= $denomP } if ($denomR -gt 0) { $recArr[0] = $nlEnv.TruePositives; $recArr[0] /= $denomR } $precPct = [Math]::Round($precArr[0] * 100, 1) $recPct = [Math]::Round($recArr[0] * 100, 1) Write-Host "" Write-Host "╔══════════════════════════════════════════════════╗" -ForegroundColor Cyan Write-Host "║ Phase 13: NL Interface - Results ║" -ForegroundColor Cyan Write-Host "╠══════════════════════════════════════════════════╣" -ForegroundColor Cyan Write-Host ("║ Baseline (random) avg reward : {0,8} ║" -f $bAvg) -ForegroundColor Gray Write-Host ("║ Trained (DQN) avg reward : {0,8} ║" -f $tAvg) -ForegroundColor Green Write-Host ("║ Improvement : {0,7}% ║" -f $improvement) -ForegroundColor Yellow Write-Host "╠══════════════════════════════════════════════════╣" -ForegroundColor Cyan Write-Host ("║ Precision (Orc+Esc correct) : {0,7}% ║" -f $precPct) -ForegroundColor Cyan Write-Host ("║ Recall (escalations caught): {0,7}% ║" -f $recPct) -ForegroundColor Cyan Write-Host "╠══════════════════════════════════════════════════╣" -ForegroundColor Cyan Write-Host "║ Agent learned to: ║" -ForegroundColor Cyan Write-Host "║ Respond simple information queries ║" -ForegroundColor White Write-Host "║ Execute clear single-agent commands ║" -ForegroundColor White Write-Host "║ Orchestrate complex multi-agent pipelines ║" -ForegroundColor White Write-Host "║ Escalate ambiguous or high-risk commands ║" -ForegroundColor White Write-Host "╚══════════════════════════════════════════════════╝" -ForegroundColor Cyan Write-Host "" return @{ Agent = $agent; Results = $results; Baseline = @{ Avg = $bAvg }; Trained = @{ Avg = $tAvg } } } # ============================================================ # TEST SUGGESTIONS # ============================================================ # 1. Run VBAF.LoadAll.ps1 (loads core DQN + all pillars) # # 2. QUICK DEMO (simulated NL intent signals) # $r = Invoke-VBAFNLInterfaceTraining -Episodes 100 -PrintEvery 10 -SimMode # # 3. FULL TRAINING (real PS host + command routing demo) # $r = Invoke-VBAFNLInterfaceTraining -Episodes 100 -PrintEvery 10 # # 4. SKIP REAL DATA PROBE # $r = Invoke-VBAFNLInterfaceTraining -Episodes 100 -PrintEvery 10 -SkipRealData # # 5. INSPECT AGENT DECISIONS # $env = [NLInterfaceEnvironment]::new() # $state = $env.Reset() # Write-Host "Confidence: $($env.IntentConfidence) Risk: $($env.RiskLevel)" # $action = $r.Agent.Act($state) # $labels = @("Respond","Execute","Orchestrate","Escalate") # Write-Host "NL Router decision: $($labels[$action])" # # 6. VIEW CONFUSION MATRIX # Write-Host "True Positives : $($env.TruePositives)" # Write-Host "False Positives: $($env.FalsePositives)" # Write-Host "True Negatives : $($env.TrueNegatives)" # Write-Host "False Negatives: $($env.FalseNegatives)" # ============================================================ Write-Host "📦 VBAF.Enterprise.NLInterface.ps1 loaded [v3.3.0 💬]" -ForegroundColor Green Write-Host " Phase 13: Natural Language Interface" -ForegroundColor Cyan Write-Host " Function : Invoke-VBAFNLInterfaceTraining" -ForegroundColor Cyan Write-Host "" Write-Host " Quick start:" -ForegroundColor Yellow Write-Host ' $r = Invoke-VBAFNLInterfaceTraining -Episodes 100 -PrintEvery 10 -SimMode' -ForegroundColor White Write-Host "" |