Set up experiments for report 1

[gianlucatruda]

UPDATE: Working on this in PR #15

Progress

• [x] Evolution step
• [x] Testing step
• [x] Use gain instead of fitness in test experiment
• [ ] Bug: individuals produce identical scores on all 5 test repeats*.
• [x] Enhancement: make evaluation functions and controllers generalised in experiments_test.py
• [x] Optional: parallelise evolution step's inner loop

WHAT: Implement two EAs that use the simulation mode “Individual evolution” of the framework to train a specialist agent.

HOW: Use 3 enemies (games) to experiment with your algorithms. For each enemy, make an independent experiment evolving an specialist agent, and for all cases use exactly the same algorithm parameters. You are allowed, but not required, to use the code of the neural network (demo_controller.py) available in the scripts of demonstration.

For the “two EAs”, you can use two different EAs, or the same EA with differences such as, the evolutionary mechanisms, parameter tuning, etc.

With both algorithms (and for each enemy) you must repeat your final experiments 10 times (independently) and your report should present the statistics based on these 10 runs.

Compare your algorithms by enemy, testing 5 times your final best solution for each of the 10 independent runs, and present the individual gain in box-plots.

[gianlucatruda]

Evolution step (pseudocode)

for ea_instance in [ea_config_1, ea_config_2]:
    for enemy in [1, 3, 5]:
        for repeat in range(10):
            # Evolve for N generations
            # Save best individual

Testing step (pseudocode)

for enemy in [1, 3, 5]:
   for ea_instance in [ea_config_1, ea_config_2]:
        for top_ten in best_individuals[enemy]:
            for individual in top_ten:
                for repeat in range(5):
                    # Test individual and average results

[gianlucatruda]

I'm honestly not sure how to interpret the assignment brief. It's so confusingly worded.