Grim Dawn Gear Optimizer

A genetic algorithm solver for finding optimal gear combinations in Grim Dawn. This project models gear optimization as a variation of the knapsack problem, maximizing a "benefit score" based on gear stats across 11 equipment slots with components and augments.

Highlights:

• Fast convergence - Typically finds optimal solutions in 6-8 seconds
• Parallel execution - 7-8x speedup on multi-core systems (enabled by default)
• Massive search space - Efficiently handles combinatorial explosions across gear/component/augment possibilities
• Lock specific items - Build around items with unique bonuses (+skills, set bonuses)
• Live progress - Real-time updates every 2 seconds with Ctrl+C support

Usage

# Run with defaults (parallel mode, auto-stop)
python3 solver.py
# Output:
# [2.0s] Best: 2148 pts | Gen 37 | 45k evals/s
# [4.0s] Best: 2159 pts | Gen 80 | 48k evals/s
# [6.0s] Best: 2159 pts | Gen 128 | 51k evals/s
# No improvement after 2s plateau. Terminating search.

# Single-core mode (slower, lower CPU usage)
python3 solver.py --single-core

# Customize time limit (default: unlimited, stops on 2s plateau)
python3 solver.py --time-limit 10

# Verbose output (show generation progress and item counts)
python3 solver.py -v

# Show only gear slots with changes (default shows all gear)
python3 solver.py --swaps-only

# Press Ctrl+C anytime to stop and see best results found so far

How It Works

The Goal

The solver finds the combination of gear that maximizes a "benefit score" by:

• Selecting one base item for each of 11 gear slots (helm, chest, shoulders, gloves, pants, boots, belt, amulet, medal, ring 1, ring 2)
• Attaching one component and one augment to each base item (when compatible options exist)
• Calculating total benefit points based on weighted stat contributions with configurable caps

The Spreadsheet (grim_items.ods)

A spreadsheet contains all available gear, components, and augments with the following structure:

• Header Row: Defines the stats available from gear (e.g., rFire for Fire Resistance)
• points_per_unit Row: Specifies the benefit score awarded for each point of a given stat
• baseline Row: Starting stat values from skills and difficulty (e.g., -50 for resistances)
• max Row: Defines the cap for benefit scoring, NOT a hard limit on gear stats
  • Example: If rFire has max of 80 and points_per_unit of 2:
    • 30 rFire from gear scores 60 points (30 × 2)
    • 50 rFire from gear scores 80 points (benefit capped at 80)
  • Gear stats can exceed the max value; excess doesn't contribute to score
• type Column: Specifies the item category (e.g., Helm, Chest, Component, Augment)
• tf Column: A '1' indicates the gear is currently equipped
• name Column: Item name (prefix with * to lock, see Locking Gear section)

Components and Augments

The game allows one "Component" and one "Augment" to be attached to each base gear piece:

• Represented in the spreadsheet with type of "Component" or "Augment"
• applies_to Column: Specifies which gear types they can attach to (e.g., "Helm,Chest" or "armor" for all armor pieces)
• The solver creates virtual items combining base + component + augment stats
• Goal: Find the best combination of all three for every slot

Multiple-Slot Items

Some gear types can be equipped in multiple slots:

• Rings use separate types (Ring 1 and Ring 2) for two independent slots
• The same ring cannot be equipped in both slots

Solver Strategy

The solver uses a Multi-Gene Genetic Algorithm as its only optimization method:

Performance:

• Adaptive termination (runs until plateau), typically 6-8 seconds
• 100% optimal on current dataset, evaluates millions of combinations per second
• Island model parallelization enabled by default (7-8x speedup on 16 cores)

Why Genetic Algorithm?

• The search space is astronomically large: each of 11 gear slots can be filled with (base × component × augment) combinations
• With typical datasets, total combinations range from billions to trillions, making brute force completely infeasible
• Multi-gene representation allows evolution of building blocks (base/component/augment) independently
• Efficiently explores trillion-scale search spaces through intelligent sampling
• Finds optimal or near-optimal solutions in seconds (typically 6-8s to convergence)
• Scales well as dataset grows

Key Features:

• Adaptive termination stops after 2s without improvement (no arbitrary timeout)
• Live progress updates every 2 seconds (score, generation, eval rate)
• Ctrl+C support for graceful interruption at any time
• Multi-gene chromosome: Each slot has 3 independent genes (base item, component, augment)
• Operators: Tournament selection, multi-gene crossover, granular mutation
• Scales population by CPU cores for better exploration (150 per core)
• Island model parallelization (enabled by default, -s for single-core): True multi-core execution with migration
  • 7-8x speedup on 16 cores (16k → 128k comparisons/sec)
  • Ring topology migration every 2 seconds (adaptive to search progress)
  • Each island evolves independently, sharing best solutions periodically

Parameters:

• Population: 150 per core (auto-calculated based on CPU count)
• Mutation rate: 15% (applies to each gene independently)
• Crossover rate: 70% (can mix genes from different parents)
• Elitism: Top 10% preserved each generation

Features

Locking Gear

Lock specific items to force them into all optimal builds (useful for items with unique bonuses like +skills that aren't captured in stat columns):

1. Open grim_items.csv (or edit grim_items.ods and reconvert)
2. Find the item you want to lock
3. Add a * prefix to its name (e.g., thread of mortality → *thread of mortality)
4. Save and run the solver

Use cases:

• Items with unique skill bonuses (e.g., +3 to all skills) not in stat columns
• Set items you want to keep for set bonuses
• Items you like aesthetically or for specific build requirements
• Testing "what's the best build around this specific item?"

How it works:

• Locked items are forced into every individual during population initialization
• The genetic algorithm never mutates locked slots
• The solver optimizes everything else around the locked items
• Components and augments can still be optimized for locked gear pieces
• The * prefix remains visible in all output displays
• At startup, the solver displays a message listing all locked items detected

Example output:

Locked items detected (1 items will be forced into all builds):
  amulet: *thread of mortality

AMULET
  Base:       *thread of mortality
  Component:  dread skull
  Augment:    boneback spine → breath of life (rov)

Note: You can lock base items, components, AND augments independently. For example, you could lock a base helm while allowing the solver to optimize its component and augment.

Output Format

The solver displays the optimal gear configuration with:

• Gear recommendations showing all slots with stats for each item (sorted by benefit value)
• Full stat comparison table (current vs recommended with benefit points)
• Execution summary (evaluations, time, benefit improvement)

Example Output:

========================================================================================================================
RECOMMENDED GEAR CONFIGURATION
========================================================================================================================

HELM
  Base:       overseers vitality  rStun(52), aAbsorb(25), hlth(891), da(78)
  Component:  runestone → leathery hide (hs)  rStun(25)
  Augment:    (none) → outcasts warding pow.  rAeth(12)

CHEST
  Base:       shrewd menhirs → impenetrable vitality  rPrc(91), hlth(947), rChs(22)
  Component:  hallowed ground → chains of oleron  oa(60), move(6)
  Augment:    (none) → outcasts warding pow.  rAeth(12)

========================================================================================================================
BASELINE STATS (from skills + difficulty)
========================================================================================================================
rPhys: -50 | rPrc: -50 | rVit: -50 | rAeth: -50 | rChs: -50 | rEle: -50 | rBld: -50 | rStun: -50

========================================================================================================================
STAT COMPARISON
========================================================================================================================
Stat      Wgt      Cur → Opt          Points Sources (optimal)
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────
dam       0.02     3805 → 3805          76   ring 1(3805)
dAll      0.33      108 → 49            16   gloves(49)
LS        7.00       11 → 13            91   belt(8), comp(5)
hlth      0.05     3940 → 4579         247   chest(947), helm(891), medal(619), gloves(344), boots(372), +2
%hlth     3.22       14 → 32           103   comp(12), boots(5), medal(5), pants(5), ring 2(5)
move      3.00       16 → 35 / 66      105   boots(19), comp(16)
+as%      5.00       15 → 24           120   gloves(17), ring 2(7)
rPhys     3.00       27 → 77 / 105     231   chest(44), amulet(33), ring 1(25), ring 2(21), helm(4)
rPrc      1.00       27 → 84 / 105      84   chest(49), amulet(33), ring 1(25), ring 2(21), comp(6)
rVit      1.00       15 → 27 / 80       27   comp(35), medal(17), aug(15)
rAeth     1.00       -2 → 20 / 80       20   aug(27), amulet(25), shoulders(18)
rChs      1.00       62 → 86 / 80       80   chest(22), ring 1(22), medal(21), ring 2(21)
rBld      1.00       59 → 74 / 80       74   pants(22), medal(20), ring 2(17), comp(15)
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────
TOTAL:  1501 → 1666 (+165 points, +11.0%)

Gear display format:

• Shows all gear slots by default (use --swaps-only to show only changes)
• Each item displays its stats sorted by benefit contribution (highest first)
• Format: item name stat(value), stat(value), ...
• Changed items show: old → new stat(value), ...
• Unchanged items show: name stat(value), ... (for verification)
• Stats truncate with +N if too many to fit on one line

Stat table column descriptions:

• Stat: Stat name from CSV header
• Wgt: Benefit points awarded per unit (shows 2 decimal places)
• Cur → Opt: Current value → Optimal value / Cap (cap shown only when present)
• Points: Total benefit points contributed by this stat in optimal build
• Sources: Items contributing to this stat (with values in parentheses)

Future Work

User Features:

• Display top-N alternative builds (diversity selection code exists but output not shown to user)
• Gear culling recommendations (identify items safe to discard based on top solutions)
• Minimum stat constraints (e.g., "must have at least 80% rAether")

Optimization Ideas:

• Seed initial population with current gear (guarantees never worse than equipped)
• Local search / hill climbing post-processing (try all single-item swaps after GA convergence)
• Smart mutation strategy (prioritize swaps that meet minimum thresholds)
• Better migration strategy (synchronous barriers, track migration success rate)
• Adaptive population sizing based on search space size