Configuration¶

DASA's behavior can be tuned through various parameters in the optimization process.

Optimization Parameters¶

These parameters are set in test.py:run_optimization():

Iterations¶

num_iterations = 2000  # Maximum optimization steps

More iterations allow finding harder-to-reach targets but increase runtime.

Learning Rate¶

initial_lr = 0.1  # Adam optimizer learning rate

Higher values: Faster convergence but may overshoot
Lower values: More stable but slower

Sigmoid Annealing¶

sigmoid_annealing_start = 0.001
sigmoid_annealing_end = 1.0

Controls how "soft" control flow decisions are during optimization:

Early iterations: Soft decisions allow gradient flow through all branches
Later iterations: Sharp decisions commit to specific paths

Temperature Annealing (Strings)¶

temp_start = 2.0   # High: exploratory sampling
temp_end = 0.1     # Low: sharp character selection

For Gumbel-Softmax string generation:

High temperature: Smooth probability distribution
Low temperature: Near-deterministic character selection

Early Stopping¶

min_loss_delta = 1e-12  # Stop if loss changes less than this

Optimization stops early when loss plateaus.

Docker Configuration¶

Environment Variables¶

Variable	Default	Description
`TARGET`	`Main`	Main class name to analyze

Volume Mounts¶

docker run --rm \
  -v "/path/to/java/files:/SUT" \      # Java source files
  -v "/path/to/verifier:/svHelpers" \  # Verifier.java location
  dasa-graph-extractor

Test Execution Parameters¶

When calling test.main():

Parameter	Type	Description
`target_file`	`str`	Path to JSON graph file
`start_nodes`	`list`	Input nodes (or `None` for auto)
`end_nodes`	`list`	Target nodes (or `None` for auto)
`auto_detect_start_end`	`bool`	Auto-detect input/target nodes
`test_dir`	`str`	Directory with .class files
`test_class`	`str`	Main class name
`use_sv_helpers`	`bool`	Include svHelpers in classpath
`return_successfull_output`	`bool`	Return stdout on success
`num_iterations`	`int`	Optimization attempts per target
`verbose`	`bool`	Print iteration progress

Initial Value Strategies¶

DASA uses smart initialization for different types:

Integers¶

# 70% chance: 2-14 bit values (common range)
# 30% chance: 15-31 bit values (edge cases)
# 5% chance: special values (0, -1, 1, MIN, MAX)

Floats¶

Similar to integers but includes floating-point ranges.

Booleans¶

# 50% true, 50% false

Characters¶

# 90% chance: printable ASCII (32-125)
# 10% chance: full Unicode range

Strings¶

# Default length: 10 characters
# Initialization: uniform distribution or biased toward constants

Timeout Configuration¶

In test.py:

# 10-minute global timeout per analysis
if datetime.now() - start_time >= timedelta(minutes=10):
    break

GraalVM Options¶

In scripts/entrypoint.sh:

native-image -ea \
  -H:Dump=:1 \                  # Dump compiler graphs
  -H:MaximumInliningSize=0 \    # Disable inlining
  -H:+UnlockExperimentalVMOptions \
  $TARGET

If graph generation fails, it retries with -O0 (no optimization).

Customizing Node Behavior¶

To modify how specific nodes behave, edit files in nodes/:

nodes/
├── calc/           # Math operations
├── java/           # Java-specific operations
├── types/          # Complex types (String, Array)
├── custom/         # Custom operations (Sigmoid)
└── BaseNode.py     # Base class for all nodes

Each node can define:

exec(): Computation logic
node_penalty: Additional loss term
controlFlowMultiplicative: Path reachability score