Genetic Material

The agent's chromosome has two parts, the head and the body. The body of an agent determines the actions and observations it will use over its lifetime, and the head determines the mutation rates of the body parts, as well as a probabilistic rule which selects an action, if the body fails to determine one.

The agent's body is a modified probabilistic finite state automata (or pFSA, see section 2.4). The modification is that each state has a particular observation associated with it, which is why the range the observation values can take is usually small. The values that the observation can return make up the alphabet for arcs coming from that particular state, and large alphabets make for overly complex pFSAs. The arcs in the pFSA have actions associated with them, along with a matching number from the state's alphabet (or observation range), and a probability.

The agent uses its pFSA body by performing one state transition per cycle. First, the observation associated with the state is made, then an arc is chosen probabilistically according to the value generated by the observation. The action associated with that arc is performed by the agent, and the current state is updated. This process is begun at an agent's start state, which is either inherited from one of its parents or chosen at random. It may happen that no arc matches the value returned by the observation. If this is the case, a random action is selected, weighted by a probability vector, and there is a small binomial chance of a random state transition. This is to prevent the agent from getting into a state where it only depends on the probability vector.

Figure 1 shows a possible deliberative agent. The state with the double circle represents the agent's start state and the text within a circle indicates an observation. Along each arc is the matching observation value, and a corresponding action. The agent keeps selecting food until it gets one good food value. As the food-check is probabilistic, the agent checks the food again from the second food-check state. If the food is good, it eats it. If it is bad, there's a 50% chance that it returns to the start state and a 50% chance that it stays in the same state. If it eats its food, the agent may choose to procreate, or walk, and be back in its original state.

  

Figure 1: Schematic diagram of an agent's pFSA brain