The forest beckons

Trollskog is a desktop city-builder strategy game for Windows due to be released in 2017.

The world of Trollskog is a dark forest, teeming with mythical creatures: some friendly, others not. Settle your towns, develop your empire, and explore deep into the heart of the forest.

If you want to follow the development of Trollskog, consider signing up for the mailing list. If you'd like more frequent updates, you can follow the game's development on twitter or facebook.

The game is currently in closed alpha testing. Find out more about the game in the About section, or sign up to join open alpha testing here.


Fireside devlog

In these balmy summer temperatures, my code flows like molasses. But as the nights get longer and cooler, development thaws.

I'm currently investigating some glitches related to saving and loading: the world coming up ever-so-slightly different each time you load it is enough to make one question one's sanity.

When I've resolved this I intend to ship a new build and send out a new wave of alpha invites. I know there are some people who have been waiting for months - you will absolutely get your invite before the game's release, and the game will be better for it. Thank you for your patience!

And a glad midsummer to you!

Summer solstice is here! Development of the game will be on hold this weekend, as there are certain rites and rituals that must be observed to guarantee a good harvest.

So we wish you a merry midsummer, and if you'd like to read more about Trollskog, check out the devlog. Meanwhile, life in the village continues...

Trade route established

A trader's caravan moves brings goods from the blue city to the orange. A profitable endeavor, but sure to attract raiders.


Sleepy times in the forest. A new unit "Idle" visual communicates to the player when their villagers are slacking off!

Pathfinding in an infinite world

One of the most important aspects in a strategy game lies in its pathfinding algorithm. This determines how the game’s entities move around in the world, which greatly affects the feel of the game.
As a player, you want the entities in the world to feel like an extension of yourself, and to watch your villagers struggle to carry out basic tasks due to a poor pathfinding algorithm is a frustrating experience.

I prototyped the pathfinding system using A* - this is the canonical pathfinding algorithm, which unfortunately isn't a viable solution in a game like Trollskog, no matter how much you dress it up, A* ends up having to explore huge areas of the world to provide search results, and we can’t be waiting around longer than 100ms for a path request without it negatively impacting the game experience.

The first serious attempt was Jump Point Search (JPS). In the same way that A* is an optimized variant of Dijkstras that assumes all edges have a non-negative length, JPS is an optimized variant of A* that assumes the traversed graph is a grid. I first encountered the algorithm in this whitepaper by Daniel Harabor and Alban Grastien.
Since the terrain in Trollskog certainly is a grid, the algorithm seemed like a good fit.

While JPS was orders of magnitude faster than A*, it wasn’t the performant silver bullet I’d hoped for, and it didn’t help at all with steering behaviors - A single entity following a path worked fine, but having a large group of entities all navigating together, bumping into each other and other entities along the way resulted in a lot of edge-cases, sometimes requiring path recalculation. I kept piling on band-aids for a while, trying to fix new strange behaviors as they arose, but eventually made the decision to try another algorithm.

The lit area indicates a route calculated by the JPS pathfinder. The yellow squares indicate jump points. Path-smoothing is applied to the route, as JPS tends to generate straight routes that take as few 45 degree turns as possible.

Some research into my problems led me to flowfields. These are great because they practically solve pathfinding and steering using the same algorithm. This GameAIPro chapter by Elijah Emerson is still the best writeup I’ve found on the subject.

Flowfields work quite differently from graph-traversal algorithms like A* and its variants. To summarize, it works by calculating a vector to the goal from each tile on the map, then each entity just has to follow the average vector of the four tiles around it to find the goal. This makes the algorithm scale very well with multiple entities;
Filling the entire world with vectors for every move command would be crazy inefficient, so instead we break the world up into sectors, and only fill the sectors that contain moving entities.

Here we see a 64x64 chunk of Trollskog’s world, divided into 8x8 sectors.

These sectors play two roles, one as a convenient bound for our vectorfield floodfill (to avoid having to fill the entire game world with vectors) and also for the first-pass hierarchical pathfinding. We can let the pathfinding system do an initial, broad-strokes search by traversing whole sectors at a time before we run a finer search on a tile-level, letting us break up a potentially large request into many smaller ones.

We need to keep these pathfinding sectors up-to-date by recalculating them, individually, whenever an entity that blocks a tile is placed or removed. And naturally, we want to generate these pathfinding regions as late in the terrain generation process as possible - An average 64x64 chunk of terrain in Trollskog contains about 2000 trees, and keeping the pathfinding regions up-to-date during the generation process while those trees are being placed would be a pointless waste of resources.

The advantages to this approach were immediately apparent. In the worst-case, A* and JPS explored most of the world before concluding that, yes, this unreachable spot surrounded by trees is, in fact, unreachable. The hierarchical flowfield approach can often rule out destinations which are not connected to any portal instantly, or at worst rule them out during the initial portal traversal stage.

Above, two entities are pathfinding between two sectors, with debug display enabled. Here is a larger group with a lower individual collision radius.

The result works a lot better than the previous (JPS) implementation, with a lot of tweakable variables (like collision radius, repelling force, etc) that I can play around with to make the end result look and feel the way I want.