There's an article in the 2017 volume of Bio-Complexity.
Here's how it starts:That's an incredibly bad start.
While it may be technically true that coevolutionary searches have changing fitness functions, Ewert and Marks have omitted the basic defining characteristic of a coevolutionary search and the reason why the fitness of a solution cannot be determined just from that solution: that coevolutionary searches are evolutionary algorithms that model two or more populations pitted against each other, and the fitness of an individual organisation is based on how well it competes against one or more members of the other populations, rather than against a fixed standard. Nor can I find any reference to this basic information elsewhere in the paper.
Ewert and Marks continue:There's a major problem here - that view applies to all evolutionary search algorithms, not just coevolutionary ones. All that is needed is a means of comparing two potential solutions and determining a winner - and this not only works just as well for multiple competing populations as it does for a single population, but it's trivial to use the same fitness function in both cases. Ewert and Marks don't seem to be aware of this. In fact they don't seem to be aware of the difference between evolutionary and coevolutionary searches at all. That can be determined from this totally false differentiation that seems to assume single-population evolutionary searches need exact fitness functions:and much more easily by noting that the reference they cite for checkers doesn't even using a coevolutionary search.*
While on their checkers cite, Ewert and Marks write this:Unfortunately for them, reading the cited paper shows that the prior knowledge is limited to how the game of checkers is played, and doesn't include anything at all about strategy or positional evaluation. Ewert and Marks are complaining that the fitness function for evolving a checkers strategy included having the evolving strategies play checkers against each other, rather than competing at tossing coins, rock-paper-scissors, mahjjong or ten-pin bowling.
Ewert and Marks have demonstrated once again that the first thing you'll learn if you read their paper is that they know almost nothing about evolutionary algorithms and simulations, have probably never written one, and (as with their article on Steiner trees) may never have even run one.
Roy
*"Survival is determined by the quality of play in a series of checkers games played against opponents from the same population."
Here's how it starts:That's an incredibly bad start.
While it may be technically true that coevolutionary searches have changing fitness functions, Ewert and Marks have omitted the basic defining characteristic of a coevolutionary search and the reason why the fitness of a solution cannot be determined just from that solution: that coevolutionary searches are evolutionary algorithms that model two or more populations pitted against each other, and the fitness of an individual organisation is based on how well it competes against one or more members of the other populations, rather than against a fixed standard. Nor can I find any reference to this basic information elsewhere in the paper.
Ewert and Marks continue:There's a major problem here - that view applies to all evolutionary search algorithms, not just coevolutionary ones. All that is needed is a means of comparing two potential solutions and determining a winner - and this not only works just as well for multiple competing populations as it does for a single population, but it's trivial to use the same fitness function in both cases. Ewert and Marks don't seem to be aware of this. In fact they don't seem to be aware of the difference between evolutionary and coevolutionary searches at all. That can be determined from this totally false differentiation that seems to assume single-population evolutionary searches need exact fitness functions:and much more easily by noting that the reference they cite for checkers doesn't even using a coevolutionary search.*
While on their checkers cite, Ewert and Marks write this:Unfortunately for them, reading the cited paper shows that the prior knowledge is limited to how the game of checkers is played, and doesn't include anything at all about strategy or positional evaluation. Ewert and Marks are complaining that the fitness function for evolving a checkers strategy included having the evolving strategies play checkers against each other, rather than competing at tossing coins, rock-paper-scissors, mahjjong or ten-pin bowling.
Ewert and Marks have demonstrated once again that the first thing you'll learn if you read their paper is that they know almost nothing about evolutionary algorithms and simulations, have probably never written one, and (as with their article on Steiner trees) may never have even run one.
Roy
*"Survival is determined by the quality of play in a series of checkers games played against opponents from the same population."
Comment