Random properties and determination

As far as the agent is concerned, nothing is random. The world is deterministic. This means that the percepts are always correct and that the actions always work.

An agent could do well in a single world merely by knowing a sequence of actions that works well in that world. But to make the simulating more interesting and challenging, there are randomization features that exist in the simulator to make it possible to define classes of worlds. The challenge is to develop agents that work in as broad classes as possible. An example of a class of worlds is:

All worlds surrounded by walls with 4×4 interior tiles, each with a probability of 0.2 of being a fatal danger. There is exactly 1 enemy and exactly 1 item. The agent starts in the southwest corner, which is the only home. The agent is facing east and has -10000 points and 1 unit of ammunition. Each action costs 1 point. The agent gets 1000 points for grabing the item and 10000 points for ending at home (where it started). The neutralization range is unlimited.

This class of worlds can be described like this in Asimulator rules:

Wall_Probability = 0.0
Danger_Probability = 0.2
Number_Of_Random_Items = 1
Number_Of_Random_Enemies = 1
Number_Of_Random_Dangers = 0
Move_Cost = 1
Turn_Cost = 1
Grab_Cost = 1
Neutralize_Cost = 1
Home_Reward = 10000
Item_Reward = 1000
Neutralize_Reward = 0
Neutralization_Range = 0
Agent_Score = -10000
Agent_Ammunition = 1
A map with the size 5×5 where the border tiles are of the type wall and known. The interior tiles are of the type random with random item and random enemy, except for the southwest interior tile, which has the global coordinates (1, 4) and is clear and known with home and random item. The agent is there facing east.