For benefits of testing AGI, is using a high-level video game description language (VGDL) gives more reliable and accurate results of general intelligence than using Arcade Learning Environment (ALE)?
Here is a description of the input to an ALE agent: Percept state: A single game screen (frame): a 2D array of 7-bit pixels, 160 pixels wide by 210 pixels high. Actions: 18 discrete actions defined by the joystick controller
Regarding VGDL, as far as I can see, the main site associated with it is gvgai.net, which is currently down. The associated API is described in this paper.
Percept state for GVGAI is more structured than for ALE, but the closest correspondence to ALE appears to be an 'Observation grid', consisting of a 2D array of sprite identifiers.
Actions: ACTION_NIL, ACTION_UP, ACTION_LEFT, ACTION_DOWN, ACTION_RIGHT and ACTION_USE (stated as 'typical' values).
Of the two, it would seem that ALE is more suitable for AGI, because of the more 'free form' nature of the input.
However, one of the issues with either of these approaches is that the set of possible actions is strongly constrained. These domains are therefore 'operationalised' - the hard task of working out what actions are possible has already been solved for the AI by the API, effectively acting as a bottleneck on the complexity of mapping from input to output.
A range of alternative game-playing frameworks are listed here and one alternative (which I personally believe is more useful for AGI purposes) is the Artificial Text Adventurer, in which (at each turn) agent is presented with natural language input describing the scene and must then output a command in natural language. Disclaimer: I am associated with this competition.