Why Soccer Players Yell: Using RoboCup to Model the Advantage of Signaling

We report an experiment where the RoboCup simulation environment was used to study the cognitive advantage provided by signals, which we view as task-specific structures generated in the environment to improve decision-making. We used the passing problem in RoboCup as our test problem and soccer-players' 'yells' of their 'passability' as the taskspecific signals. We found that yells improved accuracy– agents using the yells to decide the best player performed much better than agents computing the best pass themselves. The accuracy advantage derives from the task-specific nature of the yell, and such task-specific structures (signals) are used by organisms across species. From this, we reason that player yells are an instantiation of an implicit, evolved, and adaptive strategy, rather than an explicitly reasoned-out process. Many organisms generate stable structures in the world to reduce cognitive complexity, for themselves, for others, or both. Wood mice (Apodemus sylvaticus) distribute small objects, such as leaves or twigs, as points of reference while foraging. They do this even under laboratory conditions, using plastic discs. Such "way-marking" diminish the likelihood of losing interesting locations during foraging (Stopka & MacDonald, 2003). Red foxes (Vulpes vulpes) use urine to mark food caches they have emptied. This marking acts as a memory aid and helps them avoid unnecessary search (Henry, 1977, reported in Stopka & MacDonald, 2003). The male bower bird builds colorful bowers (nest-like structures), which are used by females to make mating decisions (Zahavi & Zahavi, 1997). Many birds advertise their desirability as mates using some form of external structure, like colorful tails, bibs etc (Bradbury & Vehrencamp, 1998). Other animals have signals that convey important information about themselves to possible mates and even predators (Zahavi & Zahavi, 1997). Such epistemic structures (Chandrasekharan & Stewart, 2004), usually termed signals, form a very important aspect of animal life across biological niches. These structures allow the organisms to hive off part of their cognitive load to the world. How much cognitive advantage do such structures provide in noisy, dynamic and adversarial environments? Where do the advantages come from? What are its components? These are the problems we address in this paper. We used the RoboCup simulation environment to study the cognitive advantage provided by epistemic structure strategy. Signaling is generally studied as communication (considered a good thing), and most approaches do not focus on the computational advantages provided by signaling. To understand the computational advantage provided by signaling, consider the peacock’s tail, the paradigmatic instance of an animal signal. The tail’s function is to allow female peacocks (peahens) to make a mating judgment, by selecting the most-healthy male (Zahavi & Zahavi, 1997). The tail reliably describes the inner state of the peacock, that it is healthy (and therefore has good genes). To see the cognitive efficiency of this mechanism, imagine the peahen having to make a mating decision without the existence of such a direct and reliable signal. The peahen will need to have a knowledge base of how the internal state, of health, can be inferred from behavioral and other cues. Let’s say “good dancing”, “lengthy chase of prey”, “long flights” (peacocks fly short distances), “tough beak” and “good claws” are cues for the health of a peacock. To arrive at a decision using these cues, first the peahen will need to “know” these cues, and that some combinations of them imply that the male is healthy. Armed with this knowledge, the female has to sample males for an extended period of time, and go through a lengthy sorting process based on the cues (rank each male on each of these cues: good, bad, okay). Then it has to compare the different results, keeping all of them in memory, to arrive at an optimal mating decision. This is a computationally intensive process. The tail allows the female peacock to shortcut all this computation, and go directly to the most-healthy male in a lot. The tail provides the peahen a single, chunked, cue, which it can compare with other similar ones perceptually (i.e. without computation) to arrive at a decision. The tail is a task-specific structure. It exists just for the peahen to make the mating decision. The other cues (like tough_beak etc.) do not exist for this purpose, they are task-neutral structures, which have to be synthesized by the pea-hen into a task-specific structure, to help with the mating decision. The tail, being a task-specific structure, allows the pea-hen to short-cut this synthesizing process. The tail 'fits' the peahen's task, and provides a standardized way of arriving at a decision, with the least amount of computation. The peacock describes its system state using its tail. Such selfdescription is one of nature’s ways of avoiding long-winded sorting and inference. The peacock example (and others above) shows that the reduction of others’ cognitive complexity using taskspecific external informational structures (what we term epistemic structures) is very common, and it can be considered one of the building blocks of nature. Note that while we develop the notion of epistemic structure using stable (or quasi-permanent) structures added to the world like markers and tails, task-specificity is a property common to all signals, including transient environment structures (like vocal signals for mating, warning etc.). In our view, transient signals are an adaptation of the basic epistemic structure theme (of adding task-specific structures to the world), to suit the highly dynamic or adversarial nature of such decision-making environments. In other words, stability is not the crucial property for being an epistemic structure/signal, task-specificity is. To understand the advantage provided by signaling, we have to understand the efficiency provided by task-specific structures. A Taxonomy of Agent-Environment Relations Even though signaling is a basic structure of cognition, it has received very little attention as a cognitive strategy. In the following section we develop a framework to understand how signaling, or the epistemic structure strategy (where the environment is changed in a way that it contributes taskspecific structures for decision-making), fits in with other agent-environment relationships. We categorize agent-world relations into four strategies. To illustrate these strategies, we use the design problem of providing disabled people access to buildings. There are four general strategies to solve this problem. Strategy 1: This involves building an all-powerful, James Bond-style vehicle that can function in all environments. It can run, jump, fly, climb spiral stairs, raise itself to high shelves, detect curbs etc. This design does not incorporate detailed environment structure into the vehicle, it is built to overcome the limitations of all environments. Strategy 2: This involves studying the vehicle's environment carefully and using that information to build the vehicle. For instance, the vehicle will take into account the existence of curbs (and them being short), stairs being non-spiral and having rails, level of elevator buttons etc. So it will have the capacity to raise itself to short curbs, climb short flight of straight stairs by making use of the rails etc. Note that the environment is not changed here. Strategy 3: This involves adding structure to the environment. For instance, building ramps and special doors so that a simple vehicle can have maximum access. This is the most elegant solution, and the most widely used one. Here structure is added to the environment, the world is “doped”, so that it contributes to the agent's task. Our analysis will focus on this approach. Strategy 4: This strategy is similar to the first, but here the environment is all-powerful instead of the vehicle. The environment becomes “smart”, and the building detects all physically handicapped people, and glides a ramp down to them, or lifts them up etc. This solution is an extreme case of strategy III, we will ignore it in the following analysis. The first strategy is similar to the centralized AI one, which ignores the structure provided by specific environments. The environment is something to be overcome, it is not considered a resource. This strategy tries to load every possible environment on to the agent, as centrally stored representations. The agent tries to map the encountered world on to this internal template structure. The second strategy is similar to the situated AI model promoted by Rodney Brooks (1991). This strategy recognizes the role of the environment as a resource, and analyses and exploits the detailed structure that exists in the environment to help the agent. Notice the environment remains unchanged, it is considered a given. The third strategy is similar to one aspect of distributed cognition, where task-specific structures are generated in the environment, allowing the agent to hive off part of the computation to the world. Kirsh (1996) terms this kind of “using the world to compute” active redesign. This strategy underlies many design techniques to minimize complexity. At the physical level, the strategy can be found in the building of roads for wheeled vehicles. Without roads, the vehicles will have a hard time, or all vehicles will need to have tank wheels. With roads, the movement is a lot easier for average vehicles. This principle is also at work in the “intelligent use of space” where people organize objects around them in a way that helps them execute their functions (Kirsh, 1995). Kitchens and personal libraries (which use locations as tags for identifying content) are instances of such use of space in cognition. Another application of task-specific structures is bar coding. Without bar coding, the checkout machine in the supermarket wou