Skip to content

Parameters and effective inputs

Documentation index · Scientific interpretation · Experiment reference

This table is generated from JobSubmissionRequest and ExperimentBudget in the current source. Regenerate with uv run --no-sync python scripts/documentation_reference.py; check drift with --check. These are application/API defaults, not an article protocol or the defaults of the separate direct mapf solve command.

One job or one expanded matrix trial

Field Default Schema constraint
grid_width 8 integer; min 1, max 64
grid_height 8 integer; min 1, max 64
obstacles "empty collection" array; max items 4096
starts "empty collection" object
goals "empty collection" object
setting "SETTING_1" SETTING_1, SETTING_2, SETTING_3, SETTING_4
name "Custom scenario" string; max characters 160
solver_id "Decentralized-HeatMap" string
scenario_id null string; max characters 100 or null
max_steps 100 integer; min 1, max 10000
timeout_sec 5.0 number; min 0.1, max 600.0 or null
negotiation_deadline_sec 60.0 number; min 0.01, max 600.0
fov_size 3 integer; min 3, max 15
broadcast_horizon null integer; min 1, max 100 or null
negotiation_horizon null integer; min 1, max 100 or null
negotiation_protocol "taop-v2" taop-v2, taop-v1, legacy-alternating-v1
negotiation_round_limit 30 integer; min 1, max 500
verification_pass_limit 5 integer; min 1, max 20
max_astar_expansions 1500 integer; min 100, max 100000
initial_tokens 10 integer; min 0, max 100
commitment_type "SC" SC, DC, ZC
random_seed 42 integer; min 0, max 4294967295
recording_level "full-trace" metrics-only, events, full-trace
heat_recording_limit 250000 integer; min 0, max 2000000
suboptimality 1.1 number; min 1.0, max 3.0

Scenario dictionaries use integer [x, y] coordinates. Starts/goals need the same IDs and at least one agent; semantic validation follows schema validation. A built-in or workspace-owned scenario_id resolves a snapshot; portable external batches should carry inline coordinates. Grid shape, starts, goals, obstacles and name are scenario data, not solver parameters.

FoV is an odd square width (3, 5, 7, 9, 11, 13 or 15); the radius is floor(width/2). Null temporal horizons are resolved to the FoV width in the effective plan. Those horizons count states, including the current position; they are not distances or seconds. The interactive GUI form caps max_steps at 500; the application/headless schema permits up to 10,000.

timeout_sec: null is accepted only for decentralized jobs. Centralized jobs require a number. The bilateral deadline resets for a new session, never for an offer. In TAOP v2 the round limit is a diagnostic checkpoint, not a hard terminal offer count. Heat values are recorded only for applicable full-trace HeatMap decisions, bounded by the declared value budget and internal record-count guard.

suboptimality is active only for solvers/profiles that support it. It is not a general guarantee for a solver with a familiar name. Native catalogs also constrain the setting and freeze executable provenance. Preview reports canonical solver identity and inactive_parameters; consult /api/v1/capabilities for controls supported by your installed catalog.

Whole experiment

Field Default Schema constraint
workers 2 integer; min 1, max 6
wall_seconds 900 number; min 1, max 86400
max_trials 10000 integer; min 1, max 100000
disk_mb 2048 integer; min 64, max 65536
threads_per_worker 1 integer; min 1, max 1
resources null object or null

The product of scenario count and matrix-axis lengths is checked against max_trials before exclusions. Fixed admission allows up to four workers; an explicit resources policy allows up to six. The GUI defaults to two workers and accepts an explicit --workers / --resource-policy configuration. Its experiment worker/policy settings must match the owning supervisor; the queue contains at most 32 pending/active jobs. Numerical-library thread caps are one per worker; they are not an OS memory guarantee.

Wall time is charged across resume. Disk limits stop work/admission when exhausted; do not treat them as a promise about the eventual compressed artifact size. The default interactive repository quota is 512 MiB; the batch runner applies its declared experiment disk budget. A GUI experiment must fit its existing workspace quota.

Optional resource policy

Field Default Schema constraint
centralized_slots 2 integer; min 1, max 6
decentralized_slots 4 integer; min 1, max 6
max_cpu_percent 90 number; max 100
free_memory_mb 1024 integer; min 0
max_owned_memory_mb 18432 integer; min 1
optimal_reserve_mb 10240 integer; min 1
bounded_reserve_mb 2048 integer; min 1
decentralized_reserve_mb 256 integer; min 1
idle_timeout_sec 300 number; min 0.1, max 86400

Set budget.resources to opt in; null retains fixed admission. These are admission estimates and thresholds, not solver limits or OS memory guarantees. See resource-aware batches for blocked-state, resume and GUI policy semantics.

Recording choice

Level Retained evidence Intended use
metrics-only Paths, validation, measured counters, run/configuration provenance; no detailed replay frames/events/local heat Broad batches and compact outcome analysis
events Replay frames and recorded telemetry events; no full local decision/heat payloads Movement and event inspection
full-trace Available local observations, plans, commitments and applicable heat records, plus replay/events Detailed explanations and debugging

Recording limits and omissions are explicit. No later export can recreate a local decision that was never recorded. Replay is a visual artifact, not an exact-resume solver checkpoint.