The proposed AM-LSTM agent demonstrated a shift in strategy.
We trained three baseline architectures against built-in heuristic bots (Rule-Based Agents or RBAs): doomz.io
DOOMZ.IO is a proposed browser-based and API-accessible benchmark that merges the high-octane mechanics of First-Person Shooters (FPS) with the resource-scarcity mechanics of "Battle Royale" and ".io" genre games. The core contribution of this paper is the definition of the DOOMZ.IO environment and the analysis of agent failure modes when subjected to the "Doom Clock"—a mechanic where the playable area shrinks as resources deplete. The proposed AM-LSTM agent demonstrated a shift in strategy
The field of Multi-Agent Reinforcement Learning (MARL) has made significant strides in zero-sum games (Chess, Go) and team-based shooters (Doom, Quake III). However, the intersection of strategic resource management and high-fidelity tactical combat remains underexplored. We introduce DOOMZ.IO , a stochastic, partially observable simulation environment designed to push the boundaries of autonomous agent capabilities. DOOMZ.IO forces agents to balance the exploration-exploitation trade-off in a rapidly shrinking spatial domain while managing a finite, decaying health resource. We demonstrate that current state-of-the-art algorithms, such as PPO and QMIX, fail to converge on optimal strategies when faced with the "Survival Paradox"—the necessity of high-risk aggression to secure longevity—often defaulting to degenerate passive strategies. We propose a novel architectural augmentation, the Adversarial Memory LSTM (AM-LSTM) , which significantly improves survival rates by predicting opponent intent based on resource scarcity cues. The field of Multi-Agent Reinforcement Learning (MARL) has