Neural Inference

uses System.Data, System.Data.Tabular;

var model := new TabularData;

// 1. Load and Parse Dataset
var csvPath := 'data/titanic.csv';
if not FileExists(csvPath) then begin
   PrintLn('Error: Titanic dataset not found at ' + csvPath);
   Exit;
end;

var lines := FileReadLines(csvPath);
var colSex, colPclass, colAge, colActual: array of Float;

for var i := 1 to lines.Length - 1 do begin
   var line := lines[i];
   if line = '' then continue;
   var inQuote := False;
   for var j := 1 to line.Length do begin
      if line[j] = '"' then inQuote := not inQuote;
      if inQuote and (line[j] = ',') then line[j] := ';';
   end;
   var fields := line.Split(',');
   if fields.Length < 6 then continue;
   
   colActual.Add(StrToFloatDef(fields[1], 0.0));
   colSex.Add(if fields[4].Contains('female') then 0.0 else 1.0);
   var pc := StrToIntDef(fields[2], 3);
   colPclass.Add(if pc = 1 then 1.0 else if pc = 2 then 0.5 else 0.0);
   colAge.Add(StrToFloatDef(fields[5], 28.0) / 80.0);
end;

model.AddColumn('Actual', colActual);
model.AddColumn('Sex', colSex);
model.AddColumn('PClass', colPclass);
model.AddColumn('Age', colAge);

PrintLn('<h3>Titanic Model Optimization</h3>');
PrintLn(Format('Searching for optimal interaction weights on %d passengers...', [colActual.Length]));

// 2. Automated Weight Search
var bestAcc := -1.0;
var bestMaleBoost := 0.0;
var bestChildBoost := 0.0;
var bestPoorFemalePenalty := 0.0;
var bestBias := 0.0;

for var i := 0 to 4 do begin
   var mb := Float(i) * 3.0; // Male 1st Class boost
   for var j := 0 to 4 do begin
      var cb := Float(j) * 3.0; // Child boost
      for var l := 0 to 4 do begin
         var pfp := Float(l) * 3.0; // Poor Female Penalty
         for var k := 0 to 10 do begin
            var bias := Float(k) * 0.2;
            
            model.EvaluateNewColumn('TmpProb', [
              1, 1, '"Sex"', -10.0, '*', 4.0, '+', 'dup0', 
                '"Sex"', '"PClass"', '*', mb, '*', '+', 
                0.2, '"Age"', '-', 0, 'max', cb, '*', '+', 
                '"Sex"', 0, '=', '"PClass"', 0, '=', '*', pfp, '*', '-', 
                0, 'relu', '+', bias, '+',
              -1, '*', 'exp', '+', '/' 
            ]);
            
            model.EvaluateNewColumn('TmpMatch', [ '"TmpProb"', 0.5, '>=', '"Actual"', '=' ]);
            var acc := (model.EvaluateAggregate('sum', ['"TmpMatch"']) / colActual.Length) * 100;
            
            if acc > bestAcc then begin
               bestAcc := acc;
               bestMaleBoost := mb;
               bestChildBoost := cb;
               bestPoorFemalePenalty := pfp;
               bestBias := bias;
            end;
            
            model.DropColumn('TmpProb');
            model.DropColumn('TmpMatch');
         end;
      end;
   end;
end;

PrintLn(Format('<b>Search Complete!</b> Found optimal weights: MaleBoost=%.1f, ChildBoost=%.1f, PoorPenalty=%.1f, Bias=%.1f', 
  [bestMaleBoost, bestChildBoost, bestPoorFemalePenalty, bestBias]));

// 3. Evaluate Final Models
model.EvaluateNewColumn('ProbA', [
  1, 1, '"Sex"', -10.0, '*', 5.0, '+', -1, '*', 'exp', '+', '/' 
]);

model.EvaluateNewColumn('ProbB', [
  1, 1, 
    '"Sex"', -10.0, '*', 4.0, '+', 
    'dup0', 
    '"Sex"', '"PClass"', '*', bestMaleBoost, '*', '+', 
    0.2, '"Age"', '-', 0, 'max', bestChildBoost, '*', '+', 
    '"Sex"', 0, '=', '"PClass"', 0, '=', '*', bestPoorFemalePenalty, '*', '-', 
    0, 'relu', '+', bestBias, '+', 
  -1, '*', 'exp', '+', '/' 
]);

model.EvaluateNewColumn('MatchA', [ '"ProbA"', 0.5, '>=', '"Actual"', '=' ]);
model.EvaluateNewColumn('MatchB', [ '"ProbB"', 0.5, '>=', '"Actual"', '=' ]);

var accA := (model.EvaluateAggregate('sum', ['"MatchA"']) / colActual.Length) * 100;
var accB := (model.EvaluateAggregate('sum', ['"MatchB"']) / colActual.Length) * 100;

PrintLn('<h4>Final Accuracy Comparison</h4>');
PrintLn(Format('<b>Baseline (Gender Only):</b> %.1f%%', [accA]));
PrintLn(Format('<b>Optimized Residual:</b> %.1f%% (Improvement: +%.1f%%)', [accB, accB - accA]));

// 4. Sample Successes
PrintLn('<br><b>Outliers correctly identified by the Deep Model:</b>');
var count := 0;
for var i := 0 to colActual.Length - 1 do begin
   if (StrToFloat(model.ColumnStrings('MatchA')[i]) = 0.0) and (StrToFloat(model.ColumnStrings('MatchB')[i]) = 1.0) then begin
      var sex := if colSex[i] = 0 then 'Female' else 'Male';
      var pc := if colPclass[i] = 1 then '1st' else if colPclass[i] = 0.5 then '2nd' else '3rd';
      var age := Integer(colAge[i] * 80);
      PrintLn(Format('Pass #%d [%s, %s, Age %d] (Survived): Corrected by Residual Model', [i+1, sex, pc, age]));
      Inc(count);
      if count >= 3 then break;
   end;
end;