AI Cleanup and Status Effects
About this tutorial
This tutorial is free and open source, and all code uses the MIT license - so you are free to do with it as you like. My hope is that you will enjoy the tutorial, and make great games!
If you enjoy this and would like me to keep writing, please consider supporting my Patreon.
In the design document, we indicated that we'd like the AI to be smarter than the average rock. We've also added quite a few AI-related systems as we've worked through the chapters and (you may have noticed) some things like consistently applying confusion effects (and occasional problems hitting a mob that just moved) have slipped through the cracks!
As we add complexity to our monsters, it would be good to get all of this straightened out, make it easier to support new features, and handle commonalities like movement consistently.
Chained Systems
Rather than try and do everything for an NPC in one system, we could break the process out into several steps. This is more typing, but has the advantage that each step is distinct, clear and does just one thing - making it much easier to debug. This is analogous to how we are handling WantsToMelee - we're indicating an intent and then handling it in its own step - which let us keep targeting and actually fighting separated.
Let's look at the steps, and see how they can be broken down:
- We determine that it is the NPC's turn.
- We check status effects - such as Confusion to determine if they can in fact go.
- The AI module for that AI type scans their surroundings, and determines if they want to move, attack or do nothing.
- Movement occurs, which updates various global statuses.
- Combat occurs, which can kill mobs or render them unable to act in the future.
Modularizing the AI
We already have quite a few AI systems, and this is just going to add more. So let's move AI into a module. Create a new folder, src/ai - this will be the new AI module. Create a mod.rs file, and put the following into it:
#![allow(unused)] fn main() { mod animal_ai_system; mod bystander_ai_system; mod monster_ai_system; pub use animal_ai_system::AnimalAI; pub use bystander_ai_system::BystanderAI; pub use monster_ai_system::MonsterAI; }
This tells it to use the other AI modules and share them all in the ai namespace. Now move animal_ai_system, bystander_ai_system and monster_ai_system from your src directory into src\ai. In the preamble to main.rs (where you have all the mod and use statements), remove the mod and use statements for these systems. Replace them with a single mod ai; line. Finally, you can cleanup run_systems to reference these systems via the ai namespace:
#![allow(unused)] fn main() { impl State { fn run_systems(&mut self) { let mut mapindex = MapIndexingSystem{}; mapindex.run_now(&self.ecs); let mut vis = VisibilitySystem{}; vis.run_now(&self.ecs); let mut mob = ai::MonsterAI{}; mob.run_now(&self.ecs); let mut animal = ai::AnimalAI{}; animal.run_now(&self.ecs); let mut bystander = ai::BystanderAI{}; bystander.run_now(&self.ecs); ... }
In your ai/X_system files you have lines that read use super::{...}. Replace the super with crate, to indicate that you want to use the components (and other types) from the parent crate.
If you cargo run now, you have exactly the same game as before - your refactor worked!
Determining Whose Turn It Is - Initiative/Energy Cost
So far, we've handles our turns in a strict but inflexible manner: the player goes, and then all the NPCs go. Back and forth, forever. This works pretty well, but it doesn't allow for much variety: you can't have something make an entity faster than the others, all actions take the same amount of time, and it would make things like haste and slow spells impossible to implement.
Many roguelike games use a variant of initiative or initiative cost to determine whose turn it is, so we'll go with something similar. We don't want to be too random, so you don't suddenly see things speed up and slow down, but we also want to be more flexible. We also want it to be a little random, so that all the NPCs don't act at the same time by default - giving basically what we have already. It would also be good to slow down wearers of heavy armor/weapons, and have users of light equipment go faster (dagger users can strike more frequently than claymore users!).
In components.rs (and registered in main.rs and saveload_system.rs), lets make a new Initiative component:
#![allow(unused)] fn main() { #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct Initiative { pub current : i32 } }
We want the player to start with an initiative score (we'll go with 0, so they always start first). In spawners.rs, we simply add it to the player function as another component for the player:
#![allow(unused)] fn main() { .with(Initiative{current: 0}) }
We also want all NPCs to start with an initiative score. So in raws/rawmaster.rs, we add it to the spawn_named_mob function as another always-present component. We'll give mobs a starting initiative of 2 - so on the first turn they will all process just after the player (we'll worry about subsequent turns later).
#![allow(unused)] fn main() { // Initiative of 2 eb = eb.with(Initiative{current: 2}); }
That adds the component, but currently it doesn't do anything at all. We're going to start by making another new component in components.rs (and registering it in main.rs and saveload_system.rs), called MyTurn:
#![allow(unused)] fn main() { #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct MyTurn {} }
The idea behind MyTurn components is that if you have the component, then it is your turn to act - and you should be included in AI/turn control (if the player has MyTurn then we wait for input). If you don't have it, then you don't get to act. We can also use it as a filter: so things like status effects can check to see if it is your turn and you are affected by a status, and they might determine that you have to skip your turn.
Now we should make a new - simple - system for handling initiative rolls. Make a new file, ai/initiative_system.rs:
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{Initiative, Position, MyTurn, Attributes, RunState}; pub struct InitiativeSystem {} impl<'a> System<'a> for InitiativeSystem { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, Initiative>, ReadStorage<'a, Position>, WriteStorage<'a, MyTurn>, Entities<'a>, WriteExpect<'a, rltk::RandomNumberGenerator>, ReadStorage<'a, Attributes>, WriteExpect<'a, RunState>, ReadExpect<'a, Entity>); fn run(&mut self, data : Self::SystemData) { let (mut initiatives, positions, mut turns, entities, mut rng, attributes, mut runstate, player) = data; if *runstate != RunState::Ticking { return; } // We'll be adding Ticking in a moment; use MonsterTurn if you want to test in the meantime // Clear any remaining MyTurn we left by mistkae turns.clear(); // Roll initiative for (entity, initiative, _pos) in (&entities, &mut initiatives, &positions).join() { initiative.current -= 1; if initiative.current < 1 { // It's my turn! turns.insert(entity, MyTurn{}).expect("Unable to insert turn"); // Re-roll initiative.current = 6 + rng.roll_dice(1, 6); // Give a bonus for quickness if let Some(attr) = attributes.get(entity) { initiative.current -= attr.quickness.bonus; } // TODO: More initiative granting boosts/penalties will go here later // If its the player, we want to go to an AwaitingInput state if entity == *player { *runstate = RunState::AwaitingInput; } } } } } }
This is pretty simple:
- We first clear out any remaining
MyTurncomponents, in case we forgot to delete one (so entities don't zoom around). - We iterate all entities that have an
Initiativecomponent (indicating they can go at all) and aPositioncomponent (which we don't use, but indicates they are on the current map layer and can act). - We subtract one from the entity's current initiative.
- If the current initiative is 0 (or less, in case we messed up!), we apply a
MyTurncomponent to them. Then we re-roll their current initiative; we're going with6 + 1d6 + Quickness Bonusfor now. Notice how we've left a comment indicating that we're going to make this more complicated later! - If it is now the player's turn, we change the global
RunStatetoAwaitingInput- it's time to process the player's instructions.
We're also checking if it is the monster's turn; we'll actually be changing that - but I didn't want the system spinning rolling initiative over and over again if we test it!
Now we need to go into mod.rs and add mod initiative_system.rs; pub use initiative_system::InitiativeSystem; pair of lines to expose it to the rest of the program. Then we open up main.rs and add it to run_systems:
#![allow(unused)] fn main() { impl State { fn run_systems(&mut self) { let mut mapindex = MapIndexingSystem{}; mapindex.run_now(&self.ecs); let mut vis = VisibilitySystem{}; vis.run_now(&self.ecs); let mut initiative = ai::InitiativeSystem{}; initiative.run_now(&self.ecs); ... }
We've added it before the various AI functions run, but after we obtain a map index and visibility - so they have up-to-date data to work with.
Adjusting the game loop to use initiative
Open up main.rs and we'll edit RunState to get rid of the PlayerTurn and MonsterTurn entries - replacing them instead with Ticking. This is going to break a lot of code - but that's ok, we're actually simplifying AND gaining functionality, which is a win by most standards! Here's the new RunState:
#![allow(unused)] fn main() { #[derive(PartialEq, Copy, Clone)] pub enum RunState { AwaitingInput, PreRun, Ticking, ShowInventory, ShowDropItem, ShowTargeting { range : i32, item : Entity}, MainMenu { menu_selection : gui::MainMenuSelection }, SaveGame, NextLevel, PreviousLevel, ShowRemoveItem, GameOver, MagicMapReveal { row : i32 }, MapGeneration, ShowCheatMenu } }
In our main loop's match function, we can delete the MonsterTurn entry completely and adjust PlayerTurn to be a more generic Ticking state:
#![allow(unused)] fn main() { RunState::Ticking => { self.run_systems(); self.ecs.maintain(); match *self.ecs.fetch::<RunState>() { RunState::AwaitingInput => newrunstate = RunState::AwaitingInput, RunState::MagicMapReveal{ .. } => newrunstate = RunState::MagicMapReveal{ row: 0 }, _ => newrunstate = RunState::Ticking } } }
You'll also want to search main.rs for PlayerTurn and MonsterTurn; a lot of states return to one of these when they are done. They now want to return to Ticking.
Likewise, in player.rs there's a lot of places we return RunState::PlayerTurn - you'll want to change all of these to Ticking.
We'll modify the hunger clock to only tick on your turn. This actually becomes more simple; we simply join on MyTurn and can remove the entire "proceed" system:
#![allow(unused)] fn main() { use specs::prelude::*; use super::{HungerClock, RunState, HungerState, SufferDamage, gamelog::GameLog, MyTurn}; pub struct HungerSystem {} impl<'a> System<'a> for HungerSystem { #[allow(clippy::type_complexity)] type SystemData = ( Entities<'a>, WriteStorage<'a, HungerClock>, ReadExpect<'a, Entity>, // The player ReadExpect<'a, RunState>, WriteStorage<'a, SufferDamage>, WriteExpect<'a, GameLog>, ReadStorage<'a, MyTurn> ); fn run(&mut self, data : Self::SystemData) { let (entities, mut hunger_clock, player_entity, runstate, mut inflict_damage, mut log, turns) = data; for (entity, mut clock, _myturn) in (&entities, &mut hunger_clock, &turns).join() { clock.duration -= 1; if clock.duration < 1 { match clock.state { HungerState::WellFed => { clock.state = HungerState::Normal; clock.duration = 200; if entity == *player_entity { log.entries.push("You are no longer well fed.".to_string()); } } HungerState::Normal => { clock.state = HungerState::Hungry; clock.duration = 200; if entity == *player_entity { log.entries.push("You are hungry.".to_string()); } } HungerState::Hungry => { clock.state = HungerState::Starving; clock.duration = 200; if entity == *player_entity { log.entries.push("You are starving!".to_string()); } } HungerState::Starving => { // Inflict damage from hunger if entity == *player_entity { log.entries.push("Your hunger pangs are getting painful! You suffer 1 hp damage.".to_string()); } SufferDamage::new_damage(&mut inflict_damage, entity, 1, false); } } } } } } }
That leaves the files in ai with errors. We'll make the bare minimum of changes to make these run for now. Delete the lines that check the game state, and add a read storage for MyTurn. Add the turn to the join, so the entity only acts if it is their turn. So in ai/animal_ai_system.rs:
#![allow(unused)] fn main() { impl<'a> System<'a> for AnimalAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteExpect<'a, Map>, ReadExpect<'a, Entity>, ReadExpect<'a, RunState>, Entities<'a>, WriteStorage<'a, Viewshed>, ReadStorage<'a, Herbivore>, ReadStorage<'a, Carnivore>, ReadStorage<'a, Item>, WriteStorage<'a, WantsToMelee>, WriteStorage<'a, EntityMoved>, WriteStorage<'a, Position>, ReadStorage<'a, MyTurn> ); fn run(&mut self, data : Self::SystemData) { let (mut map, player_entity, runstate, entities, mut viewshed, herbivore, carnivore, item, mut wants_to_melee, mut entity_moved, mut position, turns) = data; ... for (entity, mut viewshed, _herbivore, mut pos, _turn) in (&entities, &mut viewshed, &herbivore, &mut position, &turns).join() { ... for (entity, mut viewshed, _carnivore, mut pos, _turn) in (&entities, &mut viewshed, &carnivore, &mut position, &turns).join() { }
Likewise, in bystander_ai_system.rs:
#![allow(unused)] fn main() { impl<'a> System<'a> for BystanderAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteExpect<'a, Map>, ReadExpect<'a, RunState>, Entities<'a>, WriteStorage<'a, Viewshed>, ReadStorage<'a, Bystander>, WriteStorage<'a, Position>, WriteStorage<'a, EntityMoved>, WriteExpect<'a, rltk::RandomNumberGenerator>, ReadExpect<'a, Point>, WriteExpect<'a, GameLog>, WriteStorage<'a, Quips>, ReadStorage<'a, Name>, ReadStorage<'a, MyTurn>); fn run(&mut self, data : Self::SystemData) { let (mut map, runstate, entities, mut viewshed, bystander, mut position, mut entity_moved, mut rng, player_pos, mut gamelog, mut quips, names, turns) = data; for (entity, mut viewshed,_bystander,mut pos, _turn) in (&entities, &mut viewshed, &bystander, &mut position, &turns).join() { ... }
And again in monster_ai_system.rs:
#![allow(unused)] fn main() { impl<'a> System<'a> for MonsterAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteExpect<'a, Map>, ReadExpect<'a, Point>, ReadExpect<'a, Entity>, ReadExpect<'a, RunState>, Entities<'a>, WriteStorage<'a, Viewshed>, ReadStorage<'a, Monster>, WriteStorage<'a, Position>, WriteStorage<'a, WantsToMelee>, WriteStorage<'a, Confusion>, WriteExpect<'a, ParticleBuilder>, WriteStorage<'a, EntityMoved>, ReadStorage<'a, MyTurn>); fn run(&mut self, data : Self::SystemData) { let (mut map, player_pos, player_entity, runstate, entities, mut viewshed, monster, mut position, mut wants_to_melee, mut confused, mut particle_builder, mut entity_moved, turns) = data; for (entity, mut viewshed,_monster,mut pos, _turn) in (&entities, &mut viewshed, &monster, &mut position, &turns).join() { }
That takes care of the compilation errors! Now cargo run the game. It runs as before, just a little more slowly. We'll worry about performance once we have the basics going - so that's great progress, we have an initiative system!
Handling Status Effects
Right now, we check for confusion in the monster_ai_system - and actually forgot about it in bystanders, vendors and animals. Rather than copy/pasting the code everywhere, we should use this as an opportunity to create a system to handle status effect turn skipping, and clean up the other systems to benefit. Make a new file, ai/turn_status.rs:
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, Confusion, RunState}; pub struct TurnStatusSystem {} impl<'a> System<'a> for TurnStatusSystem { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, WriteStorage<'a, Confusion>, Entities<'a>, ReadExpect<'a, RunState>); fn run(&mut self, data : Self::SystemData) { let (mut turns, mut confusion, entities, runstate) = data; if *runstate != RunState::Ticking { return; } let mut not_my_turn : Vec<Entity> = Vec::new(); let mut not_confused : Vec<Entity> = Vec::new(); for (entity, _turn, confused) in (&entities, &mut turns, &mut confusion).join() { confused.turns -= 1; if confused.turns < 1 { not_confused.push(entity); } else { not_my_turn.push(entity); } } for e in not_my_turn { turns.remove(e); } for e in not_confused { confusion.remove(e); } } } }
This is pretty simple: it iterates everyone who is confused, and decrements their turn counter. If they are still confused, it takes away MyTurn. If they have recovered, it takes away Confusion. You need to add a mod and pub use statement for it in ai/mod.rs, and add it to your run_systems function in main.rs:
#![allow(unused)] fn main() { let mut initiative = ai::InitiativeSystem{}; initiative.run_now(&self.ecs); let mut turnstatus = ai::TurnStatusSystem{}; turnstatus.run_now(&self.ecs); }
This shows the new pattern we are using: systems do one thing, and can remove MyTurn to prevent future execution. You can also go into monster_ai_system and remove everything relating to confusion.
Quipping NPCs
Remember when we added bandits, we gave them some commentary to say for flavor? You may have noticed that they aren't actually speaking! That's because we handled quipping in the bystander AI - rather than as a general concept. Let's move the quipping into its own system. Make a new file, ai/quipping.rs:
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{gamelog::GameLog, Quips, Name, MyTurn, Viewshed}; pub struct QuipSystem {} impl<'a> System<'a> for QuipSystem { #[allow(clippy::type_complexity)] type SystemData = ( WriteExpect<'a, GameLog>, WriteStorage<'a, Quips>, ReadStorage<'a, Name>, ReadStorage<'a, MyTurn>, ReadExpect<'a, rltk::Point>, ReadStorage<'a, Viewshed>, WriteExpect<'a, rltk::RandomNumberGenerator>); fn run(&mut self, data : Self::SystemData) { let (mut gamelog, mut quips, names, turns, player_pos, viewsheds, mut rng) = data; for (quip, name, viewshed, _turn) in (&mut quips, &names, &viewsheds, &turns).join() { if !quip.available.is_empty() && viewshed.visible_tiles.contains(&player_pos) && rng.roll_dice(1,6)==1 { let quip_index = if quip.available.len() == 1 { 0 } else { (rng.roll_dice(1, quip.available.len() as i32)-1) as usize }; gamelog.entries.push( format!("{} says \"{}\"", name.name, quip.available[quip_index]) ); quip.available.remove(quip_index); } } } } }
This is basically the quipping code from bystander_ai_system, so we don't really need to go over it in too much detail. You do want to add it into run_systems in main.rs so it functions (and add a mod and pub use statement in ai/mod.rs):
#![allow(unused)] fn main() { turnstatus.run_now(&self.ecs); let mut quipper = ai::QuipSystem{}; quipper.run_now(&self.ecs); }
Also go into bystander_ai_system.rs and remove all the quip code! It shortens it a lot, and if you cargo run now, Bandits can insult you. In fact, any NPC can be given quip lines now - and will merrily say things to you. Once again, we've made the system smaller and gained functionality. Another win!
Making AIs appear to think
Currently, we have a separate system for every type of AI - and wind up duplicating some code as a result. We also have some pretty unrealistic things going on: monsters remain completely static until they can see you, and forget all about you the moment you round a corner. Villagers move like random drunks, even when sober. Wolves hunt down deer - but again, only when they are visible. You can achieve a considerable increase in apparent AI intelligence (it's still quite dumb!) by giving NPCs goals - and having the goal last more than one turn. You can then switch the type-based decision making to be goal-based; helping the NPC achieve whatever it is that they want in life.
Let's take a moment to consider what our NPCs really want in life:
- Deer and other herbivores really want to eat grass, be left alone, and run away from things likely to kill them (which is everything, really; not a great place to be on the food chain).
- Monsters want to guard the dungeon, kill players, and otherwise lead a peaceful life.
- Wolves (and other carnivores) want to snack on players and herbivores.
- Gelatinous Cubes aren't really known for thinking much!
- Villagers really want to go about their daily lives, occasionally saying things to passing players.
- Vendors want to stay in their shops and sell you things in a future chapter update!
That doesn't really take into account transient objectives; an injured monster might want to get away from the fight, a monster might want to consider picking up the glowing Longsword of Doom that happens to be right next to them, and so on. It's a good start, though.
We can actually boil a lot of this down to a "state machine". You've seen those before: RunState makes the whole game a state, and each of the UI boxes returns a current state. In this case, we'll let an NPC have a state - which represents what they are trying to do right now, and if they have achieved it yet. We should be able to describe an AI's goals in terms of tags in the json raw file, and implement smaller sub-systems to make the AI behave somewhat believably.
Determining how an AI feels about other entities
A lot of decisions facing AIs revolve around: who is that, and how do I feel about them? If they are an enemy, I should either attack or flee (depending upon my personality). If I feel neutral towards them, then I don't really care about their presence. If I like them, I may even want to stick close by! Entering every entity's feelings towards every other entity would be a huge data-entry chore - every time you added an entity, you'd need to go and add them to every single other entity (and remember to remove/edit them everywhere if you remove them/want to change them). That's not a great idea!
Like a lot of games, we can resolve this with a simple faction system. NPCs (and the player) are members of a faction. The faction has feelings towards other factions (including a default). We can then do a simple faction lookup to see how an NPC feels about a potential target. We can also include faction information in the user interface, to help players understand what's going on.
We'll start with a faction table in spawns.json. Here's a first draft:
"faction_table" : [
{ "name" : "Player", "responses": { }},
{ "name" : "Mindless", "responses": { "Default" : "attack" } },
{ "name" : "Townsfolk", "responses" : { "Default" : "ignore" } },
{ "name" : "Bandits", "responses" : { "Default" : "attack" } },
{ "name" : "Cave Goblins", "responses" : { "Default" : "attack" } },
{ "name" : "Carnivores", "responses" : { "Default" : "attack" } },
{ "name" : "Herbivores", "responses" : { "Default" : "flee" } }
],
We'd also need to add an entry to each NPC, e.g.: "faction" : "Bandit".
To make this work, we need to create a new component to store faction membership. As always, it needs registration in main.rs and saveload_system.rs as well as definition in components.rs:
#![allow(unused)] fn main() { #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct Faction { pub name : String } }
Let's start to use this by opening up spawner.rs and modifying the player function to always add the player to the "Player" faction:
#![allow(unused)] fn main() { .with(Faction{name : "Player".to_string() }) }
Now we need to load the faction table while we load the rest of the raw data. We'll make a new file, raws/faction_structs.rs to hold this information. Our goal is to mirror what we came up with for the JSON:
#![allow(unused)] fn main() { use serde::{Deserialize}; use std::collections::HashMap; #[derive(Deserialize, Debug)] pub struct FactionInfo { pub name : String, pub responses : HashMap<String, String> } }
In turn, we add it to the Raws structure in raws/mod.rs:
#![allow(unused)] fn main() { #[derive(Deserialize, Debug)] pub struct Raws { pub items : Vec<Item>, pub mobs : Vec<Mob>, pub props : Vec<Prop>, pub spawn_table : Vec<SpawnTableEntry>, pub loot_tables : Vec<LootTable>, pub faction_table : Vec<FactionInfo> } }
We also need to add it to the raw constructor in raws/rawmaster.rs:
#![allow(unused)] fn main() { impl RawMaster { pub fn empty() -> RawMaster { RawMaster { raws : Raws{ items: Vec::new(), mobs: Vec::new(), props: Vec::new(), spawn_table: Vec::new(), loot_tables: Vec::new(), faction_table : Vec::new(), }, item_index : HashMap::new(), mob_index : HashMap::new(), prop_index : HashMap::new(), loot_index : HashMap::new() } } }
We'll also want to add some indexing into Raws. We need a better way to represent reactions than a string, so lets add an enum to faction_structs.rs first:
#![allow(unused)] fn main() { #[derive(PartialEq, Eq, Hash, Copy, Clone)] pub enum Reaction { Ignore, Attack, Flee } }
Now we add an index for reactions to RawMaster:
#![allow(unused)] fn main() { pub struct RawMaster { raws : Raws, item_index : HashMap<String, usize>, mob_index : HashMap<String, usize>, prop_index : HashMap<String, usize>, loot_index : HashMap<String, usize>, faction_index : HashMap<String, HashMap<String, Reaction>> } }
Also add it to the RawMaster constructor as faction_index : HashMap::new(). Finally, we'll setup the index - open the load function and add this at the end:
#![allow(unused)] fn main() { for faction in self.raws.faction_table.iter() { let mut reactions : HashMap<String, Reaction> = HashMap::new(); for other in faction.responses.iter() { reactions.insert( other.0.clone(), match other.1.as_str() { "ignore" => Reaction::Ignore, "flee" => Reaction::Flee, _ => Reaction::Attack } ); } self.faction_index.insert(faction.name.clone(), reactions); } }
This iterates through all of the factions, and then through their reactions to other factions - building a HashMap of how they respond to each faction. These are then stored in the faction_index table.
So that loads the raw faction information, we still have to turn it into something readily usable in-game. We should also add a faction option to the mob_structs.rs:
#![allow(unused)] fn main() { #[derive(Deserialize, Debug)] pub struct Mob { pub name : String, pub renderable : Option<Renderable>, pub blocks_tile : bool, pub vision_range : i32, pub ai : String, pub quips : Option<Vec<String>>, pub attributes : MobAttributes, pub skills : Option<HashMap<String, i32>>, pub level : Option<i32>, pub hp : Option<i32>, pub mana : Option<i32>, pub equipped : Option<Vec<String>>, pub natural : Option<MobNatural>, pub loot_table : Option<String>, pub light : Option<MobLight>, pub faction : Option<String> } }
And in spawn_named_mob, add the component. If there isn't one, we'll automatically apply "mindless" to the mob:
#![allow(unused)] fn main() { if let Some(faction) = &mob_template.faction { eb = eb.with(Faction{ name: faction.clone() }); } else { eb = eb.with(Faction{ name : "Mindless".to_string() }) } }
Now in rawmaster.rs, we'll add one more function: to query the factions table to obtain a reaction about a faction:
#![allow(unused)] fn main() { pub fn faction_reaction(my_faction : &str, their_faction : &str, raws : &RawMaster) -> Reaction { if raws.faction_index.contains_key(my_faction) { let mf = &raws.faction_index[my_faction]; if mf.contains_key(their_faction) { return mf[their_faction]; } else if mf.contains_key("Default") { return mf["Default"]; } else { return Reaction::Ignore; } } Reaction::Ignore } }
So, given the name of my_faction and the other entity's faction (their_faction), we can query the faction table and return a reaction. We default to Ignore, if there isn't one (which shouldn't happen, since we default to Mindless).
Common AI task: handling adjacent entities
Pretty much every AI needs to know how to handle an adjacent entity. It might be an enemy (to attack or run away from), someone to ignore, etc. - but it needs to be handled. Rather than handling it separately in every AI module, lets build a common system to handle it. Let's make a new file, ai/adjacent_ai_system.rs (and add a mod and pub use entry for it in ai/mod.rs like the others):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, Faction, Position, Map, raws::Reaction, WantsToMelee}; pub struct AdjacentAI {} impl<'a> System<'a> for AdjacentAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, ReadStorage<'a, Faction>, ReadStorage<'a, Position>, ReadExpect<'a, Map>, WriteStorage<'a, WantsToMelee>, Entities<'a>, ReadExpect<'a, Entity> ); fn run(&mut self, data : Self::SystemData) { let (mut turns, factions, positions, map, mut want_melee, entities, player) = data; let mut turn_done : Vec<Entity> = Vec::new(); for (entity, _turn, my_faction, pos) in (&entities, &turns, &factions, &positions).join() { if entity != *player { let mut reactions : Vec<(Entity, Reaction)> = Vec::new(); let idx = map.xy_idx(pos.x, pos.y); let w = map.width; let h = map.height; // Add possible reactions to adjacents for each direction if pos.x > 0 { evaluate(idx-1, &map, &factions, &my_faction.name, &mut reactions); } if pos.x < w-1 { evaluate(idx+1, &map, &factions, &my_faction.name, &mut reactions); } if pos.y > 0 { evaluate(idx-w as usize, &map, &factions, &my_faction.name, &mut reactions); } if pos.y < h-1 { evaluate(idx+w as usize, &map, &factions, &my_faction.name, &mut reactions); } if pos.y > 0 && pos.x > 0 { evaluate((idx-w as usize)-1, &map, &factions, &my_faction.name, &mut reactions); } if pos.y > 0 && pos.x < w-1 { evaluate((idx-w as usize)+1, &map, &factions, &my_faction.name, &mut reactions); } if pos.y < h-1 && pos.x > 0 { evaluate((idx+w as usize)-1, &map, &factions, &my_faction.name, &mut reactions); } if pos.y < h-1 && pos.x < w-1 { evaluate((idx+w as usize)+1, &map, &factions, &my_faction.name, &mut reactions); } let mut done = false; for reaction in reactions.iter() { if let Reaction::Attack = reaction.1 { want_melee.insert(entity, WantsToMelee{ target: reaction.0 }).expect("Error inserting melee"); done = true; } } if done { turn_done.push(entity); } } } // Remove turn marker for those that are done for done in turn_done.iter() { turns.remove(*done); } } } fn evaluate(idx : usize, map : &Map, factions : &ReadStorage<Faction>, my_faction : &str, reactions : &mut Vec<(Entity, Reaction)>) { for other_entity in map.tile_content[idx].iter() { if let Some(faction) = factions.get(*other_entity) { reactions.push(( *other_entity, crate::raws::faction_reaction(my_faction, &faction.name, &crate::raws::RAWS.lock().unwrap()) )); } } } }
This system works as follows:
- We query all entities with a faction, a position, and a turn and make sure we aren't modifying the player's behavior by checking the entity with the player entity resource.
- We query the map for all adjacent tiles, recording reactions to neighboring entities.
- We iterate the resulting reactions, if it is an
Attackreaction - we cancel their turn and initiate aWantsToMeleeresult.
To actually use this system, add it into run_systems in main.rs, before the MonsterAI:
#![allow(unused)] fn main() { let mut adjacent = ai::AdjacentAI{}; adjacent.run_now(&self.ecs); }
If you cargo run the game now, pandemonium erupts! Everyone is in the "mindless" faction, and as a result is hostile to everyone else! This is actually a great demo of how our engine can perform; despite combat going on from all quarters, it runs pretty well:
Restoring peace in the town
It's also not at all what we had in mind for a peaceful starting town. It might work for a zombie apocalypse, but that's best left to Cataclysm: Dark Days Ahead (an excellent game, by the way)! Fortunately, we can restore peace to the town by adding a "faction" : "Townsfolk" line to all of the town NPCs. Here's the barkeep as an example; you need to do the same for all of the towns-people:
{
"name" : "Barkeep",
"renderable": {
"glyph" : "☻",
"fg" : "#EE82EE",
"bg" : "#000000",
"order" : 1
},
"blocks_tile" : true,
"vision_range" : 4,
"ai" : "vendor",
"attributes" : {
"intelligence" : 13
},
"skills" : {
"Melee" : 2
},
"equipped" : [ "Cudgel", "Cloth Tunic", "Cloth Pants", "Slippers" ],
"faction" : "Townsfolk"
},
Once you've put those in, you can cargo run - and have peace in our time! Well, almost: if you watch the combat log, the rats lay into one another with a vengeance. Again, not quite what we intended. Open up spawns.json and lets add a faction for rats - and have them ignore one another. We'll add ignoring one another to a few other factions, too - so bandits aren't slaying one another for no reason:
"faction_table" : [
{ "name" : "Player", "responses": { }},
{ "name" : "Mindless", "responses": { "Default" : "attack" } },
{ "name" : "Townsfolk", "responses" : { "Default" : "ignore" } },
{ "name" : "Bandits", "responses" : { "Default" : "attack", "Bandits" : "ignore" } },
{ "name" : "Cave Goblins", "responses" : { "Default" : "attack", "Cave Goblins" : "ignore" } },
{ "name" : "Carnivores", "responses" : { "Default" : "attack", "Carnivores" : "ignore" } },
{ "name" : "Herbivores", "responses" : { "Default" : "flee", "Herbivores" : "ignore" } },
{ "name" : "Hungry Rodents", "responses": { "Default" : "attack", "Hungry Rodents" : "ignore" }}
],
Also, add the Rat to the Hungry Rodents faction:
{
"name" : "Rat",
"renderable": {
"glyph" : "r",
"fg" : "#FF0000",
"bg" : "#000000",
"order" : 1
},
"blocks_tile" : true,
"vision_range" : 8,
"ai" : "melee",
"attributes" : {
"Might" : 3,
"Fitness" : 3
},
"skills" : {
"Melee" : -1,
"Defense" : -1
},
"natural" : {
"armor_class" : 11,
"attacks" : [
{ "name" : "bite", "hit_bonus" : 0, "damage" : "1d4" }
]
},
"faction" : "Hungry Rodents"
},
cargo run now, and you'll see that the rats are leaving each other alone.
Responding to more distant entities
Responding to those next to you is a great first step, and actually helps with processing time (since adjacent enemies are processed without a costly search of the entire viewshed) - but if there isn't an adjacent enemy, the AI needs to look for a more distant one. If one is spotted that needs a reaction, we need some components to indicate intent. In components.rs (and registered in main.rs and saveload_system.rs):
#![allow(unused)] fn main() { #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct WantsToApproach { pub idx : i32 } #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct WantsToFlee { pub indices : Vec<usize> } }
These are intended to indicate what the AI would like to do: either approach a tile (an enemy), or flee from a list of enemy tiles.
We'll make another new system, ai/visible_ai_system.rs (and add it to mod and pub use in ai/mod.rs):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, Faction, Position, Map, raws::Reaction, Viewshed, WantsToFlee, WantsToApproach}; pub struct VisibleAI {} impl<'a> System<'a> for VisibleAI { #[allow(clippy::type_complexity)] type SystemData = ( ReadStorage<'a, MyTurn>, ReadStorage<'a, Faction>, ReadStorage<'a, Position>, ReadExpect<'a, Map>, WriteStorage<'a, WantsToApproach>, WriteStorage<'a, WantsToFlee>, Entities<'a>, ReadExpect<'a, Entity>, ReadStorage<'a, Viewshed> ); fn run(&mut self, data : Self::SystemData) { let (turns, factions, positions, map, mut want_approach, mut want_flee, entities, player, viewsheds) = data; for (entity, _turn, my_faction, pos, viewshed) in (&entities, &turns, &factions, &positions, &viewsheds).join() { if entity != *player { let my_idx = map.xy_idx(pos.x, pos.y); let mut reactions : Vec<(usize, Reaction)> = Vec::new(); let mut flee : Vec<usize> = Vec::new(); for visible_tile in viewshed.visible_tiles.iter() { let idx = map.xy_idx(visible_tile.x, visible_tile.y); if my_idx != idx { evaluate(idx, &map, &factions, &my_faction.name, &mut reactions); } } let mut done = false; for reaction in reactions.iter() { match reaction.1 { Reaction::Attack => { want_approach.insert(entity, WantsToApproach{ idx: reaction.0 as i32 }).expect("Unable to insert"); done = true; } Reaction::Flee => { flee.push(reaction.0); } _ => {} } } if !done && !flee.is_empty() { want_flee.insert(entity, WantsToFlee{ indices : flee }).expect("Unable to insert"); } } } } } fn evaluate(idx : usize, map : &Map, factions : &ReadStorage<Faction>, my_faction : &str, reactions : &mut Vec<(usize, Reaction)>) { for other_entity in map.tile_content[idx].iter() { if let Some(faction) = factions.get(*other_entity) { reactions.push(( idx, crate::raws::faction_reaction(my_faction, &faction.name, &crate::raws::RAWS.lock().unwrap()) )); } } } }
Remember that this won't run at all if we're already dealing with an adjacent enemy - so there's no need to worry about assigning melee. It also doesn't do anything - it triggers intent for other systems/services. So we don't have to worry about ending the turn. It simply scans every visible tile, and evaluates the available reactions to the tile's content. If it sees something it would like to attack, it set a WantsToApproach component. If it sees things from which it should flee, it populates a WantsToFlee structure.
You'll want to add this into run_systems in main.rs also, after the adjacency check:
#![allow(unused)] fn main() { let mut visible = ai::VisibleAI{}; visible.run_now(&self.ecs); }
Approaching
Now that we're flagging a desire to approach a tile (for whatever reason; currently because the occupant deserves a whacking), we can write a very simple system to handle this. Make a new file, ai/approach_ai_system.rs (and mod/pub use it in ai/mod.rs):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, WantsToApproach, Position, Map, Viewshed, EntityMoved}; pub struct ApproachAI {} impl<'a> System<'a> for ApproachAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, WriteStorage<'a, WantsToApproach>, WriteStorage<'a, Position>, WriteExpect<'a, Map>, WriteStorage<'a, Viewshed>, WriteStorage<'a, EntityMoved>, Entities<'a> ); fn run(&mut self, data : Self::SystemData) { let (mut turns, mut want_approach, mut positions, mut map, mut viewsheds, mut entity_moved, entities) = data; let mut turn_done : Vec<Entity> = Vec::new(); for (entity, mut pos, approach, mut viewshed, _myturn) in (&entities, &mut positions, &want_approach, &mut viewsheds, &turns).join() { turn_done.push(entity); let path = rltk::a_star_search( map.xy_idx(pos.x, pos.y) as i32, map.xy_idx(approach.idx % map.width, approach.idx / map.width) as i32, &mut *map ); if path.success && path.steps.len()>1 { let mut idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = false; pos.x = path.steps[1] as i32 % map.width; pos.y = path.steps[1] as i32 / map.width; entity_moved.insert(entity, EntityMoved{}).expect("Unable to insert marker"); idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = true; viewshed.dirty = true; } } want_approach.clear(); // Remove turn marker for those that are done for done in turn_done.iter() { turns.remove(*done); } } } }
This is basically the same as the approach code from MonsterAI, but it applies to all approach requests - to any target. It also removes MyTurn when done, and removes all approach requests. Add it to run_systems in main.rs, after the distant AI handler:
#![allow(unused)] fn main() { let mut approach = ai::ApproachAI{}; approach.run_now(&self.ecs); }
Fleeing
We'll also want to implement a system for fleeing, mostly based on the fleeing code from our Animal AI. Make a new file, flee_ai_system.rs (and remember mod and pub use in ai/mod.rs):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, WantsToFlee, Position, Map, Viewshed, EntityMoved}; pub struct FleeAI {} impl<'a> System<'a> for FleeAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, WriteStorage<'a, WantsToFlee>, WriteStorage<'a, Position>, WriteExpect<'a, Map>, WriteStorage<'a, Viewshed>, WriteStorage<'a, EntityMoved>, Entities<'a> ); fn run(&mut self, data : Self::SystemData) { let (mut turns, mut want_flee, mut positions, mut map, mut viewsheds, mut entity_moved, entities) = data; let mut turn_done : Vec<Entity> = Vec::new(); for (entity, mut pos, flee, mut viewshed, _myturn) in (&entities, &mut positions, &want_flee, &mut viewsheds, &turns).join() { turn_done.push(entity); let my_idx = map.xy_idx(pos.x, pos.y); map.populate_blocked(); let flee_map = rltk::DijkstraMap::new(map.width as usize, map.height as usize, &flee.indices, &*map, 100.0); let flee_target = rltk::DijkstraMap::find_highest_exit(&flee_map, my_idx, &*map); if let Some(flee_target) = flee_target { if !map.blocked[flee_target as usize] { map.blocked[my_idx] = false; map.blocked[flee_target as usize] = true; viewshed.dirty = true; pos.x = flee_target as i32 % map.width; pos.y = flee_target as i32 / map.width; entity_moved.insert(entity, EntityMoved{}).expect("Unable to insert marker"); } } } want_flee.clear(); // Remove turn marker for those that are done for done in turn_done.iter() { turns.remove(*done); } } } }
We also need to register in in run_systems (main.rs), after the approach system:
#![allow(unused)] fn main() { let mut flee = ai::FleeAI{}; flee.run_now(&self.ecs); }
For added effect, lets make Townsfolk run away from potentially hostile entities. In spawns.json:
{ "name" : "Townsfolk", "responses" : { "Default" : "flee", "Player" : "ignore", "Townsfolk" : "ignore" } },
If you cargo run and play now, monsters will approach and attack - and cowards will flee from hostiles.
Cleaning up
We're now performing the minimum AI performed by MonsterAI and much of the carnivore and herbivore handling in our generic systems, as well as giving townsfolk more intelligence than before! If you look at MonsterAI - there's nothing left that isn't performed already! So we can delete ai/monster_ai_system.rs, and remove it from run_systems (in main.rs) altogether! Once deleted, you should cargo run to see if the game is unchanged - it should be!
Likewise, the fleeing and approaching of ai/animal_ai_system.rs is now redundant. You can actually delete this system, too!
It would be a good idea to make sure that all NPCs have a faction (except for Gelatinous Cubes, who actually are mindless) now. You can check out the source code of spawns.json to see the changes: it's pretty obvious, everything now has a faction.
The remaining AI: Bystanders
So the remaining distinct AI module is the bystander, and they are doing just the one thing: moving randomly. This is actually a behavior that would fit well for deer, too (rather than just standing around). It would be nice if townsfolk showed slightly more intelligence, too.
Let's think about how our AI now works:
- Initiative determines if it's an NPC's turn.
- Status can take that away, depending upon effects being experienced.
- Adjacency determines immediate responses to nearby entities.
- Vision determines responses to slightly less nearby entities.
- Per-AI systems determine what the entity does now.
We could replace per-AI systems with a more generic set of "move options". These would govern what an NPC does if none of the other systems have caused it to act. Now let's think about how how we'd like townsfolk and others to move:
- Vendors stay in their shop.
- Patrons should stay in the shop they are patronizing.
- Drunk should stumble around at random. Deer should probably also move randomly, it just makes sense.
- Regular townsfolk should move between buildings, acting like they have a plan.
- Guards could patrol (we don't have any guards, but they would make sense). It might be nice for other monster types to patrol rather than staying static, also. Maybe bandits should roam the forest in search of victims.
- Hostiles should chase their target beyond visual range, but with some chance of escape.
Making a movement mode component
Let's make a new component (in components.rs, and registered in main.rs and saveload_system.rs) to capture movement mode. We'll start with the easy ones: static (not going anywhere), and random (wandering like a fool!). Note that you don't need to register the enum - just the component:
#![allow(unused)] fn main() { #[derive(Debug, Serialize, Deserialize, Clone, Eq, PartialEq, Hash)] pub enum Movement { Static, Random } #[derive(Component, Debug, Serialize, Deserialize, Clone)] pub struct MoveMode { pub mode : Movement } }
Now we'll open up raws/mob_structs.rs and edit it to capture a movement mode - and no longer provide an AI tag (since this will let us do-away with them completely):
#![allow(unused)] fn main() { #[derive(Deserialize, Debug)] pub struct Mob { pub name : String, pub renderable : Option<Renderable>, pub blocks_tile : bool, pub vision_range : i32, pub movement : String, pub quips : Option<Vec<String>>, pub attributes : MobAttributes, pub skills : Option<HashMap<String, i32>>, pub level : Option<i32>, pub hp : Option<i32>, pub mana : Option<i32>, pub equipped : Option<Vec<String>>, pub natural : Option<MobNatural>, pub loot_table : Option<String>, pub light : Option<MobLight>, pub faction : Option<String> } }
(We renamed ai to movement). This breaks a chunk of rawmaster; open up the spawn_named_mob function and replace the AI tag selection with:
#![allow(unused)] fn main() { match mob_template.movement.as_ref() { "random" => eb = eb.with(MoveMode{ mode: Movement::Random }), _ => eb = eb.with(MoveMode{ mode: Movement::Static }) } }
Now, we need a new system to handle "default" (i.e. we've tried everything else) movement. Make a new file, ai/default_move_system.rs (don't forget to mod and pub use it in ai/mod.rs!):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, MoveMode, Movement, Position, Map, Viewshed, EntityMoved}; pub struct DefaultMoveAI {} impl<'a> System<'a> for DefaultMoveAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, ReadStorage<'a, MoveMode>, WriteStorage<'a, Position>, WriteExpect<'a, Map>, WriteStorage<'a, Viewshed>, WriteStorage<'a, EntityMoved>, WriteExpect<'a, rltk::RandomNumberGenerator>, Entities<'a> ); fn run(&mut self, data : Self::SystemData) { let (mut turns, move_mode, mut positions, mut map, mut viewsheds, mut entity_moved, mut rng, entities) = data; let mut turn_done : Vec<Entity> = Vec::new(); for (entity, mut pos, mode, mut viewshed, _myturn) in (&entities, &mut positions, &move_mode, &mut viewsheds, &turns).join() { turn_done.push(entity); match mode.mode { Movement::Static => {}, Movement::Random => { let mut x = pos.x; let mut y = pos.y; let move_roll = rng.roll_dice(1, 5); match move_roll { 1 => x -= 1, 2 => x += 1, 3 => y -= 1, 4 => y += 1, _ => {} } if x > 0 && x < map.width-1 && y > 0 && y < map.height-1 { let dest_idx = map.xy_idx(x, y); if !map.blocked[dest_idx] { let idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = false; pos.x = x; pos.y = y; entity_moved.insert(entity, EntityMoved{}).expect("Unable to insert marker"); map.blocked[dest_idx] = true; viewshed.dirty = true; } } } } } // Remove turn marker for those that are done for done in turn_done.iter() { turns.remove(*done); } } } }
Now open up main.rs, find run_systems and replace the call to BystanderAI with DefaultMoveAI:
#![allow(unused)] fn main() { let mut defaultmove = ai::DefaultMoveAI{}; defaultmove.run_now(&self.ecs); }
Finally, we need to open up spawns.json and replace all references to ai= in mobs with movement=. Choose static for all of them except for Patrons, herbivores and Drunks.
If you cargo run now, you'll see that everyone stands around - except for the random ones, who wander aimlessly.
One last thing for this segment: go ahead and delete the bystander_ai_system.rs file, and all references to it. We don't need it anymore!
Adding in waypoint-based movement
We mentioned that we'd like townsfolk to mill about, but not randomly. Open up components.rs, and add a mode to Movement:
#![allow(unused)] fn main() { #[derive(Debug, Serialize, Deserialize, Clone, Eq, PartialEq, Hash)] pub enum Movement { Static, Random, RandomWaypoint{ path : Option<Vec<usize>> } } }
Notice that we're using Rust's feature that enum is really a union in other languages, to add in an optional path for random movement. This represents where the AI is trying to go - or None if there isn't a current target (either because they just started, or they got there). We're hoping to not run an expensive A-Star search every turn, so we'll store the path - and keep following it until it is invalid.
Now in rawmaster.rs, we'll add it to the list of movement modes:
#![allow(unused)] fn main() { match mob_template.movement.as_ref() { "random" => eb = eb.with(MoveMode{ mode: Movement::Random }), "random_waypoint" => eb = eb.with(MoveMode{ mode: Movement::RandomWaypoint{ path: None } }), _ => eb = eb.with(MoveMode{ mode: Movement::Static }) } }
And in default_move_system.rs, we can add in the actual movement logic:
#![allow(unused)] fn main() { Movement::RandomWaypoint{path} => { if let Some(path) = path { // We have a target - go there let mut idx = map.xy_idx(pos.x, pos.y); if path.len()>1 { if !map.blocked[path[1] as usize] { map.blocked[idx] = false; pos.x = path[1] % map.width; pos.y = path[1] / map.width; entity_moved.insert(entity, EntityMoved{}).expect("Unable to insert marker"); idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = true; viewshed.dirty = true; path.remove(0); // Remove the first step in the path } // Otherwise we wait a turn to see if the path clears up } else { mode.mode = Movement::RandomWaypoint{ path : None }; } } else { let target_x = rng.roll_dice(1, map.width-2); let target_y = rng.roll_dice(1, map.height-2); let idx = map.xy_idx(target_x, target_y); if tile_walkable(map.tiles[idx]) { let path = rltk::a_star_search( map.xy_idx(pos.x, pos.y) as i32, map.xy_idx(target_x, target_y) as i32, &mut *map ); if path.success && path.steps.len()>1 { mode.mode = Movement::RandomWaypoint{ path: Some(path.steps) }; } } } }
This is a bit convoluted, so let's walk through it:
- We match on
RandomWaypointand capturepathas a variable (to access it inside the enum). - If a path exists:
- If it has more than one entry.
- If the next step isn't blocked.
- Actually perform the move by following the path.
- Remove the first entry from the path, so we keep following it.
- Wait a turn, the path may clear up
- If the next step isn't blocked.
- Give up and set no path.
- If it has more than one entry.
- If the path doesn't exist:
- Pick a random location.
- If the random location is walkable, path to it.
- If the path succeeded, store it as the AI's
path. - Otherwise, leave with no path - knowing we'll be back next turn to try another one.
If you cargo run now (and set some AI types in spawns.json to random_waypoint), you'll see that villagers now act like they have a plan - they move along paths. Because A-Star respects our movement costs, they even automatically prefer paths and roads! It looks much more realistic now.
Chasing the target
Our other stated goal is that once an AI starts to chase a target, it shouldn't give up just because it lost line-of-sight. On the other hand, it shouldn't have an omniscient view of the map and perfectly track its target either! It also needs to not be the default action - but should occur before defaults if it is an option.
We can accomplish this by creating a new component (in components.rs, remembering to register in main.rs and saveload_system.rs):
#![allow(unused)] fn main() { #[derive(Component, Debug, ConvertSaveload, Clone)] pub struct Chasing { pub target : Entity } }
Unfortunately, we're storing an Entity - so we need some extra boilerplate to make the serialization system happy:
rust
Now we can modify our visible_ai_system.rs file to add a Chasing component whenever it wants to chase after a target. There's a lot of small changes, so I've included the whole file:
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, Faction, Position, Map, raws::Reaction, Viewshed, WantsToFlee, WantsToApproach, Chasing}; pub struct VisibleAI {} impl<'a> System<'a> for VisibleAI { #[allow(clippy::type_complexity)] type SystemData = ( ReadStorage<'a, MyTurn>, ReadStorage<'a, Faction>, ReadStorage<'a, Position>, ReadExpect<'a, Map>, WriteStorage<'a, WantsToApproach>, WriteStorage<'a, WantsToFlee>, Entities<'a>, ReadExpect<'a, Entity>, ReadStorage<'a, Viewshed>, WriteStorage<'a, Chasing> ); fn run(&mut self, data : Self::SystemData) { let (turns, factions, positions, map, mut want_approach, mut want_flee, entities, player, viewsheds, mut chasing) = data; for (entity, _turn, my_faction, pos, viewshed) in (&entities, &turns, &factions, &positions, &viewsheds).join() { if entity != *player { let my_idx = map.xy_idx(pos.x, pos.y); let mut reactions : Vec<(usize, Reaction, Entity)> = Vec::new(); let mut flee : Vec<usize> = Vec::new(); for visible_tile in viewshed.visible_tiles.iter() { let idx = map.xy_idx(visible_tile.x, visible_tile.y); if my_idx != idx { evaluate(idx, &map, &factions, &my_faction.name, &mut reactions); } } let mut done = false; for reaction in reactions.iter() { match reaction.1 { Reaction::Attack => { want_approach.insert(entity, WantsToApproach{ idx: reaction.0 as i32 }).expect("Unable to insert"); chasing.insert(entity, Chasing{ target: reaction.2}).expect("Unable to insert"); done = true; } Reaction::Flee => { flee.push(reaction.0); } _ => {} } } if !done && !flee.is_empty() { want_flee.insert(entity, WantsToFlee{ indices : flee }).expect("Unable to insert"); } } } } } fn evaluate(idx : usize, map : &Map, factions : &ReadStorage<Faction>, my_faction : &str, reactions : &mut Vec<(usize, Reaction, Entity)>) { for other_entity in map.tile_content[idx].iter() { if let Some(faction) = factions.get(*other_entity) { reactions.push(( idx, crate::raws::faction_reaction(my_faction, &faction.name, &crate::raws::RAWS.lock().unwrap()), *other_entity )); } } } }
That's a great start: when going after an NPC, we'll automatically start chasing them. Now, lets make a new system to handle chasing; create ai/chase_ai_system.rs (and mod, pub use in ai/mod.rs):
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{MyTurn, Chasing, Position, Map, Viewshed, EntityMoved}; use std::collections::HashMap; pub struct ChaseAI {} impl<'a> System<'a> for ChaseAI { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, MyTurn>, WriteStorage<'a, Chasing>, WriteStorage<'a, Position>, WriteExpect<'a, Map>, WriteStorage<'a, Viewshed>, WriteStorage<'a, EntityMoved>, Entities<'a> ); fn run(&mut self, data : Self::SystemData) { let (mut turns, mut chasing, mut positions, mut map, mut viewsheds, mut entity_moved, entities) = data; let mut targets : HashMap<Entity, (i32, i32)> = HashMap::new(); let mut end_chase : Vec<Entity> = Vec::new(); for (entity, _turn, chasing) in (&entities, &turns, &chasing).join() { let target_pos = positions.get(chasing.target); if let Some(target_pos) = target_pos { targets.insert(entity, (target_pos.x, target_pos.y)); } else { end_chase.push(entity); } } for done in end_chase.iter() { chasing.remove(*done); } end_chase.clear(); let mut turn_done : Vec<Entity> = Vec::new(); for (entity, mut pos, _chase, mut viewshed, _myturn) in (&entities, &mut positions, &chasing, &mut viewsheds, &turns).join() { turn_done.push(entity); let target_pos = targets[&entity]; let path = rltk::a_star_search( map.xy_idx(pos.x, pos.y) as i32, map.xy_idx(target_pos.0, target_pos.1) as i32, &mut *map ); if path.success && path.steps.len()>1 && path.steps.len()<15 { let mut idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = false; pos.x = path.steps[1] as i32 % map.width; pos.y = path.steps[1] as i32 / map.width; entity_moved.insert(entity, EntityMoved{}).expect("Unable to insert marker"); idx = map.xy_idx(pos.x, pos.y); map.blocked[idx] = true; viewshed.dirty = true; turn_done.push(entity); } else { end_chase.push(entity); } } for done in end_chase.iter() { chasing.remove(*done); } for done in turn_done.iter() { turns.remove(*done); } } } }
This system ended up being more complicated than I hoped, becuase the borrow checker really didn't want me reaching into Position storage twice. So we ended up with the following:
- We iterate all entities that have a
Chasingcomponent, as well as a turn. We look to see if their target is valid, and if it is - we store it in a temporary HashMap. This gets around needing to look insidePositiontwice. If it isn't valid, we remove the component. - We iterate everyone who is still chasing, and path to their target. If the path succeeds, then it follows the path. If if doesn't, we remove the chasing component.
- We remove everyone who took a turn from the
MyTurnlist.
Add it into run_systems before the default movement system:
#![allow(unused)] fn main() { let mut approach = ai::ApproachAI{}; approach.run_now(&self.ecs); }
Removing per-AI tags
We're no longer using the Bystander, Monster, Carnivore, Herbivore and Vendor tags! Open up components.rs and delete them. You'll also need to delete their registration in main.rs and saveload_system.rs. Once they are gone, you will still see errors in player.rs; why? We used to use these tags to determine if we should attack or trade-places with an NPC. We can replace the failing code in try_move_player quite easily. First, remove the references to these components from your using statements. Then replace these two lines:
#![allow(unused)] fn main() { let bystanders = ecs.read_storage::<Bystander>(); let vendors = ecs.read_storage::<Vendor>(); }
with:
#![allow(unused)] fn main() { let factions = ecs.read_storage::<Faction>(); }
Then we replace the tag check with:
#![allow(unused)] fn main() { for potential_target in map.tile_content[destination_idx].iter() { let mut hostile = true; if combat_stats.get(*potential_target).is_some() { if let Some(faction) = factions.get(*potential_target) { let reaction = crate::raws::faction_reaction( &faction.name, "Player", &crate::raws::RAWS.lock().unwrap() ); if reaction != Reaction::Attack { hostile = false; } } } if !hostile { // Note that we want to move the bystander }
Notice that we're using the faction system we made earlier! There's one more fix to player.rs - deciding if we can heal because of nearby monsters. It's basically the same change - we check if an entity is hostile, and if it is it prohibits healing (because you are nervous/on-edge!):
#![allow(unused)] fn main() { fn skip_turn(ecs: &mut World) -> RunState { let player_entity = ecs.fetch::<Entity>(); let viewshed_components = ecs.read_storage::<Viewshed>(); let factions = ecs.read_storage::<Faction>(); let worldmap_resource = ecs.fetch::<Map>(); let mut can_heal = true; let viewshed = viewshed_components.get(*player_entity).unwrap(); for tile in viewshed.visible_tiles.iter() { let idx = worldmap_resource.xy_idx(tile.x, tile.y); for entity_id in worldmap_resource.tile_content[idx].iter() { let faction = factions.get(*entity_id); match faction { None => {} Some(faction) => { let reaction = crate::raws::faction_reaction( &faction.name, "Player", &crate::raws::RAWS.lock().unwrap() ); if reaction == Reaction::Attack { can_heal = false; } } } } } ... }
Distance Culling AI
We're currently spending a lot of CPU cycles on events far from the player. Performance is still ok, but this is sub-optimal for two reasons:
- We may just run around finding dead people if the factions are fighting while we are far away. It tells a better story to arrive as something is happening rather than just finding the aftermath.
- We don't want to waste our precious CPU cycles!
Let's open up initiative_system.rs and modify it to check the distance to the player, and not have a turn if they are far away:
#![allow(unused)] fn main() { use specs::prelude::*; use crate::{Initiative, Position, MyTurn, Attributes, RunState}; pub struct InitiativeSystem {} impl<'a> System<'a> for InitiativeSystem { #[allow(clippy::type_complexity)] type SystemData = ( WriteStorage<'a, Initiative>, ReadStorage<'a, Position>, WriteStorage<'a, MyTurn>, Entities<'a>, WriteExpect<'a, rltk::RandomNumberGenerator>, ReadStorage<'a, Attributes>, WriteExpect<'a, RunState>, ReadExpect<'a, Entity>, ReadExpect<'a, rltk::Point>); fn run(&mut self, data : Self::SystemData) { let (mut initiatives, positions, mut turns, entities, mut rng, attributes, mut runstate, player, player_pos) = data; if *runstate != RunState::Ticking { return; } // Clear any remaining MyTurn we left by mistkae turns.clear(); // Roll initiative for (entity, initiative, pos) in (&entities, &mut initiatives, &positions).join() { initiative.current -= 1; if initiative.current < 1 { let mut myturn = true; // Re-roll initiative.current = 6 + rng.roll_dice(1, 6); // Give a bonus for quickness if let Some(attr) = attributes.get(entity) { initiative.current -= attr.quickness.bonus; } // TODO: More initiative granting boosts/penalties will go here later // If its the player, we want to go to an AwaitingInput state if entity == *player { *runstate = RunState::AwaitingInput; } else { let distance = rltk::DistanceAlg::Pythagoras.distance2d(*player_pos, rltk::Point::new(pos.x, pos.y)); if distance > 20.0 { myturn = false; } } // It's my turn! if myturn { turns.insert(entity, MyTurn{}).expect("Unable to insert turn"); } } } } } }
Fixing Performance
You may have noticed a performance drop while we worked through this chapter. We've added a lot of functionality, so the systems seemed like the culprit - but they aren't! Our systems are actually running at a really good speed (one advantage of doing one thing per system: your CPU cache is very happy!). If you'd like to prove it, do a debug build, fire up a profiler (I use Very Sleepy on Windows) and attach it to the game!
The culprit is actually initiative. Not every entity is moving on the same tick anymore, so it's taking more cycles through the main loop to get to the player's turn. This is a small slowdown, but noticeable. Fortunately, you can fix it with a quick change to the main loop in main.rs:
#![allow(unused)] fn main() { RunState::Ticking => { while newrunstate == RunState::Ticking { self.run_systems(); self.ecs.maintain(); match *self.ecs.fetch::<RunState>() { RunState::AwaitingInput => newrunstate = RunState::AwaitingInput, RunState::MagicMapReveal{ .. } => newrunstate = RunState::MagicMapReveal{ row: 0 }, _ => newrunstate = RunState::Ticking } } } }
This runs all initiative cycles until it's the player's turn. It brings the game back up to full speed.
Wrap-Up
This has been a long chapter, for which I apologize - but it's been a really productive one! Instead of standing around or roaming completely randomly, AI now operates in layers - deciding first on adjacent targets, then visible targets, and then a default action. It can even hunt you down. This has gone a long way to make the AI feel smarter.
If you cargo run now, you can enjoy a much richer world!
...
The source code for this chapter may be found here
Run this chapter's example with web assembly, in your browser (WebGL2 required)
Copyright (C) 2019, Herbert Wolverson.