AMISTAD Lab - Survival Advantages of Intention Perception

Introduction

Greetings! We are the Intention Perception team working directly with Professor George Montañez. Our team consists of Amani Maina-Kilaas, Cynthia Hom, Kevin Ginta, and Cindy Lay. Our research aims to show that there exist scenarios in which the ability of an agent to perceive intention provides a measurable survival advantage. We show this by developing and running simulations to explore scenarios where this knowledge may be especially beneficial.

Experiment 1: Predators and Prey

In our first experiment, we create a predator and prey scenario, in which the predators try to eat the prey, and the prey try to avoid predators and eat food. In this scenario, we give the prey three different tiers of awareness. The lowest tier is no predator awareness—the prey can only see food. The next tier is proximity awareness—they are aware of the location of nearby predators and can try to avoid them accordingly. The highest tier—the intention awareness tier—grants prey the ability to know when a specific predator is targeting them (as opposed to another nearby prey) and focus on running away. We then compare the length of the typical lifespan of these awareness tiers under varying initial conditions, such as speed and sight distance.

Our results strongly suggest that the intention tier (light blue) does in fact provide a significant advantage over the other two, as it's often several times the average lifespan of the proximity tier (medium blue). Note: click on the photos for higher quality.

Looking at the “Prey Status over Time graph,” we also see that when prey are required to eat once every 2000 time steps, prey with intention perception are far more likely to starve than the other two, but nevertheless they generally survive much longer.

If graphs aren’t your thing, then here are some visual runs of these scenarios (a.k.a. anecdotal evidence)! The predators are the large red spheres (the red lines indicate the prey they are targeting), the prey are the small blue spheres, and the food are the even smaller green cubes.

Tier 1 - Proximity + Intention: 8/20 Alive at 12000 Steps (Max)

Tier 2 - Proximity: 0/20 Alive at 1572 Steps

Tier 3 - Unaware: 0/20 Alive at 827 Steps

Experiment 2: Two-Player Game

In our second experiment, we only have two agents: a “hero” and an “adversary.” Our scenario is inspired by a scene in the show Star Wars Rebels where our protagonists have to decide whether to keep hiding from an enemy or to strike first. Just like in the show, our adversary is generally stronger than the hero and thus avoiding conflict is in the hero’s best interest—if the hero is lucky, the adversary will pass without ever finding the hero. However, attacking the adversary first provides an element of surprise and gives the hero a temporary edge in battle. So the main question remains: Should the hero stay hidden or should the hero attack?

We test several strategies for the hero: NEVER (never attack), ALWAYS (always attack), RANDOM (randomly decide when to start attacking), RETALIATE (attack only when the adversary strikes first), and INTENTION (attack once discovered but before the adversary has the chance to strike). The last option involves intention perception, as the hero has knowledge of whether the adversary intends to strike. The simulation is carried out over 10 time steps. At each step, the agents are either hiding/searching, attacking, or waiting for a short time before they can attack again.

In this experiment, we vary Pk,a, the probability that the adversary’s attack kills the hero; Pk,h, the probability that the hero’s attack kills the adversary;  Pk,s, the boost in probability of killing the adversary gained from the hero using the element of surprise; Pd, the probability that the adversary detects the hero on a given time step; and cycle, the timer between attacks. The results show that in a majority of conditions, the INTENTION strategy fares the best—only occasionally is it beaten out by the ALWAYS strategy. However, if we take adversary survival into account, considering a least-harm approach or imagining situations where an attack may have other consequences, then there is an even stronger argument for the INTENTION strategy.

Once again, we have created some simulation visuals for the graph averse. Here is an example run of three of the strategies, ordered by overall performance:

INTENTION

ALWAYS

RETALIATE

                                  

Experiment 3: The Gopher Trap

Our third experiment takes a new approach to intention perception. Different from the first two experiments, where intention was perceived from direct interaction with an agent, here, intention is perceived based on artifacts that have been left behind. 

To model this scenario, we use the analogy of a gopher deciding whether or not to enter a trap. Each trap has food at its center and contains a variety of trap components. However, the presence of trap pieces does not mean the environment purposefully targets the gopher; odds are that it’s perfectly safe to enter and eat the food. The gopher has to examine its environment and reason as to whether it was set up—whether it was actually intended to harm the gopher. 

Traps are a product of several pieces: a door, likely some wires, and a laser gun that we dubbed an arrow. If everything is properly connected, with matching rotations and thicknesses, the door will send a “pulse” through the wires to the arrow, which should then fire a laser and “zap” the gopher. Because of these requirements, a randomly generated trap layout with pieces haphazardly strewn around is unlikely to pose a threat. Traps that are designed to harm the gopher are much more likely to be “efficient,” meaning that each wire or arrow cell should have a coherent connection. To provide some knowledge of intention, the gopher can assess the ratio of coherent connections to trap pieces and compare this value to the expected distribution should the trap be randomly generated. If it is statistically surprising to stumble upon such an efficient randomly-generated trap, the gopher will reject that hypothesis, assume it was intended for the gopher, and choose to not enter the trap. 

If the gopher does decide to enter the trap and gets zapped, the strength of the laser is dependent on the thickness of the arrow, with the wide-thickness arrows having the highest chance of stopping the gopher. It seems like it would be best to never enter the trap, but if the gopher doesn’t eat within a certain number of traps, the gopher will starve. The gopher’s survival comes down to luck and its own judgement.

In summary, we’ve found that our presumption is correct. The gopher provided with intention perception tends to survive for much longer than the gopher without. Here are survival graphs where we vary how long a gopher can survive without food and the probability of encountering a real trap. In the second graph, we add a “cautious” gopher which blindly mimics the behavior of the intention gopher and demonstrates that the benefit from intention is not due to a mere increase in caution.

And here we have two stack plots, showing how gophers with and without intention compare during simulations of 50 traps. Like with experiment 1, we see that gophers with intention perception are more likely to starve than their counterparts, but still have the best survival rates:

And of course, here come the videos:

Gopher with Intention (Zapped After 18 Traps)

Gopher without Intention (Starved After 14 Traps)

Future Plans

We are currently running more simulations for Experiment 1 on Harvey Mudd’s servers. Once we have the data from the servers, we will be able to start analyzing the results and write up a definitive conclusion. Our current data is looking very promising and we are excited to be able to share our findings with everyone!

TEAM

Lab Director: Prof. George Montañez

Amani Maina-Kilaas: amainakilaas@hmc.edu

Cynthia Hom: chom@hmc.edu

Kevin Ginta: kevin.ginta@biola.edu

Cindy Lay: clay22@cmc.edu