Genetic algorithms require a general structure

[006fix]

A general structure has to be agreed for the genetic algorithms. This may vary a little based on the algorithm in question, but a general approach is likely to be generally useful. I propose the following logic:

In any given situation, the algorithm will be presented with one of the following options : upgrade main building, upgrade warehouse, upgrade granary, upgrade field (4 types).

In order to make a decision, information must be presented about these datapoints. I propose the data points initially provided are as follows:

• main building level
• warehouse level
• granary level
• average field level (by type, 4 datapoints)
• current resources stored (by type, 4 datapoints)
• turn counter

In order to use this information to make a decision, it must be used to generate a utiilty function for a given output building.

We can therefore suggest that, given there are 4 fields, and 3 buildings, the genetic algorithm consists of 7 related sets of chromosomes, henceforth referred to as a chromosome with 7 genes. Each gene has identical structure, and consists of multipliers for the 12 aforementioned datapoints.

Therefore, in the simplest possible version, each individual in the population will contain no less than 84 numbers.

There are two further extensions that could be made to this. For example, consider modification of datapoint X - the above mentioned method suggests that answer = y * x. However, it is also possible to apply either answer = (y * x)^z or answer = y * x^z. Or, indeeed, answer = (y * x^z)^a.. This would generate a chromosome with either 168, or 252 numbers.

A secondary extension would consist of modification of the output. Currently, the output of any individual gene could be considered to be (x1y1 + x2y2 + x3y3....), however multipliers between any of the 12 datapoints (aka (z1(x1y1 + x5y5)+(x2y2 + x3y3)....) are possible, as are exponent multipliers. This would add a wide number of possible varieties (not calculated to limit search space required).

[006fix]

For each gene within the chromosome, the function contained within will be used to calculate the utility function of its paired type, aka gene 2 may be used to derive the utility function of building a warehouse.

This can then be used in one of two ways, which itself can be used in one of two ways. We'll refer to these as 0 and 1, and A and B.

0 = Once utility functions are derived, select the option with the highest utility value, and build it (it can be noted that this does not specify say, which of the 4-6 fields associated with a given type are to be upgraded. This will be done by simply selecting the one with the lowest level)

1 = Once utility functions are derived, these shall be used as probability seeding. For example, consider there are two options available, with utility functions of 9 and 1 respectively (higher=better). This shall be used to seed a random selection between the two, with respective probabilities of 90% and 10% respectively.

A = Utility functions can be calculated with respect to all currently possible opportunities, and used only with respect to these. For example, the currently available options could be warehouse, granary, crop field. Utilities are calculated for these 3, and then pursuant to either the method detailed in 0 or 1, the preferential output could be chosen.

B = Utility functions calculated with respect to all theoretically possible opportunities, and used with respect to all of these. Using the above example, utility calculations could suggest that the most preferential option is to produce a wood field. This is not currently available, but will be eventually. This can then be used to generate a sleep command for the AI, waiting until the requested building/field is available for upgrade (although at this point, it is likely that recalculation would occur, which could change the results.

Any permutation of 0 and 1 and A and B (aka 0B or 1A) is possible

[006fix]

Elitism to preserve X% of the best solutions from the previous generations should also be used

[GregoryPye]

A couple of thoughts @006fix

1. this is a production lens, and reads like it would apply to a single village at a time, but in general actors think about consumption as well as production, and for all their holdings. Aka, 'what am I short of / need'.

2. worth thinking about the rate of iteration of the algorithms. If fast enough then you could allow for something more radical which is near enough random maths on the data, with survivors maths then used for further iteration, some of which is small (to help local optimisation) and some of which is larger (to help break out of local minima). And, maybe, blending the best survivors (which is harder on that random maths as it may not be compatible). I gear the model you have may not allow for enough non-linear type maths / true variation (also possible I am misreading it).

3. neural net may be a good model for 2, as it allows most types of maths in the end, and easier to blend since you just add the nets and then later weight pruning 'genetics' deal with cruft that builds up. If doing that then you might want inputs that relate to general needs and inputs that relate to a specific village, then do villages in turn, rather than inputs for all villages .... That way the input matrix remains stable size

4. You can @GregoryPye mention me on here and it will ping me :-)