There's been quite a bit of effort put into procedural platformer level design in recent years. Despite the progress made, it is commonly thought that game developers are reaching the limitations of what an artifical intelligence can create. Hoffstein wrote that "platformer level design cannot duplicated by any method other than careful design, calibration and testing." However, Spelunky (by Derek Yu) and Cloudberry Kingdom (by Pwnee Studios) clearly challenge this notion. There are countless examples of games that utilise procedurally generated environments to enhance gameplay. Hoffstein may well have fallen foul of Clarke's First Law: "When [a distinguished scientist] states that something is impossible, he is very probably wrong."
Spelunky avoids the limitations of its generation algorithm by having an entire deformable environment. If there is a problem with the orientation of obstacles in the environment, the player has enough agency within the game world to simply circumvent that problem. It could be argued that Spelunky isn't a "true platformer" in the same way that Minecraft isn't a "true RPG." Both Spelunky and Minecraft rely on emergent gameplay to drive a player-led narrative. Whilst Spelunky has been warmly recieved, it doesn't contain the core jump-a-thon mechanics that define traditional platformer gaming.
Cloudberry Kingdom, on the other hand, takes a less flexible approach to its environment; an approach that arguably leverages more artifical intelligence. In a nod to Mario ROM hacks, especially ones deliberately designed to be highly challenging such as Kaizo Mario World, Cloudberry Kingdom creates levels that exhibit complex interactions between dynamic hazards, bottomless pits and moving platforms. This, clearly, is an example of robust artifical intelligence design that utilises traditional platformer mechanics. There's a good video on their Kickstarter page that shows how detailed the game's levels can be.
Fig 1. A "challenging" end-game scenario.
How does such a system work? Platformers boil down to two basic activities: running and jumping. So long as both running and jumping follow deterministic rules, you can successfully model a hypothetical series of actions that is guaranteed to get the player from the start to the finish. In a way, you are generating a series of key presses that solve a time-dependant puzzle. Even the more complex platform abilities: wall-jumping, spin-jumping and butt-pounding can all be modelled and implemented within a system. Cloudberry Kingdom's generator simulates action after action as a string of events that must be executed at the right point in time in order to proceed. It's not dissimilar to the underlying mechanics of games such as Guitar Hero.
It's important to note that Cloudberry Kingdom generates levels purely by pre-defined rules rather than pre-defined rooms. Rogue Legacy (by Cellar Door Games) is an example of a platformer that places pre-defined rooms next to each to create a rich environment - the location of enemies is largely random but the physical environment does come in chunks. Both games feature procedural generation but to different degrees.
Cloudberry Kingdom warranted a 68% approval rating on MetaCritic. As with many games on the fringes of the games industry, it appeals to a niche audience of enthusiasts. 68% is somewhat weaker than Rogue Legacy, a game that scored 85% on MetaCritic. But why? They're both difficult platformers that combine an element of procedural generation with tradtional mechanics. The Destructoid review does a good job of covering the main flaws of Cloudberry Kingdom: unlikable music, paper-thin storyline and some uninteresting character abilities are all cited. However, none of these flaws are a direct product of the level generation algorithm, nor are they particularly damaging issues.
However, the Edge Online review claims that the levels are "cold, sterile gauntlets, straight off the randomised production line." This succinct, and damning, statement reveals two key issues:
Fig 2. Common experiences come in many forms.
1: Lack of a common experience
It is highly unlikely that two players will ever experience the same level or event. Rogue Legacy and, the structurally similar, The Binding of Isaac use pre-designed items, power-ups and bosses to create a sense of continuity and shared experience between players. These two games are eminately discussable: conversations that begin with questions such as "What's the best tactic for Khdir?" or a statement like "The Bible is so handy for spike rooms!" are indispensible tools for creating positive emotion.
2: One level, one solution
Due to the puzzle-like nature of the game, there is necessarily only a single solution for each, especially at higher levels of play. For players unfamiliar with the hyper-difficult platformer genre, this is very discouraging. Cloudberry Kingdom, in its pursuit of precision platforming, loses a great deal of character and accessibility without multiple paths and extraneous dead-ends.
Both these issues are a shortfall of the game design but not necessarily of the technology. Whilst the design decisions made are consistent with the initial scope of the game (to create an "artifical intelligence capable of doing the impossible: designing insane Mario levels"), they are design decisions that inevitably limited the appeal and scope of the game. It would seem Hoffstein's challenge and conventional wisdom has not been bested yet. The technology behind procedural generation is developing fast, very fast, with even indie developers able to create artificial intelligences that can deliver a significant challenge to experienced players. Whilst Cloudberry Kingdom has some design limitations, its entirely procedural approach to level design is novel and demands greater experimentation.