UC Berkeley

This dissertation examines a number of scenarios in multi-agent foraging. The first scenario examined looks at a group of agents seeking to forage from resources that have a known location, but unknown quality. Communication among the agents is limited by range, and resource sites are arranged so they are farther apart than the communication range of the agents. Agents seek to maximize their expected foraging rate over a given time horizon. Initially, agents must choose which site they are going to investigate first. After inspecting a site, the agents must choose to either continue foraging at the site they are currently at or to investigate one of the other sites. Agents can communicate their decisions to one another as well as their information about other food sites when they are within communication range. The optimal way of making decisions is shown mathematically for the case of two resource sites and multiple agents under certain assumptions as well as three resource sites and only one agent.

Another scenario examined is one where discrete food pieces are generated on a grid and agents must compete with one another to forage the food pieces. Food pieces are generated from two distinct resource sites. The first case examined is when the other agents are visible and the food pieces are visible to the agent. In this case the agent moves to consume a piece of food if it can do so before any other agents. If it is not moving to consume a piece of food, then it moves to maximize its probability of consuming a piece of food in the future. Another case is examined when the other agents are not visible. In this case an agent must decide whether it thinks it can consume a food piece before other agents can. The agent also wants to move in order to maximize the probability of consuming future food particles, but in this case must make decisions based on how food particles are disappearing rather than directly observing other agents.

The most complicated scenario examined is one where the other agents are visible to one another, but unconsumed food pieces are not visible. The only food locations known to an agent are the pieces that it consumes. Agents search for pieces of food based on probability maps that they generate. Probability maps are generated based on the agent's information it has about food it has consumed. The agents also make use of observations of the other agents in order to search for new food particles.