publication croisée depuis : https://lemmy.pierre-couy.fr/post/2152233

I've been playing with deep reinforcement learning for a while. I originally started with a simple DQN, added all improvements from the Rainbow paper, and finally changed C51 for a quantile regression (and plan to swap it for an Implicit Quantile Network).

After implementing C51 (which was my first time with distributional RL) I started playing with policies that take advantage of the learned distributions : By independently taking N samples from each action-value distribution, scoring actions by averaging the samples, and picking the greedy action with respect to these scores, I was able to make the agent learn faster than similar agents using only NoisyNets or an epsilon-greedy policy (I'm still using NoisyNet, this is done on top of it). In the limiting cases, N=1 is just Thompson Sampling and N=+Infinity is just a plain greedy policy.

Finding an optimal value for N proved to be a challenge, so I decided to pick a random value for it at the start of each episode (N = 2**rng.uniform(8,12) for a QR-DQN with 32 quantiles/action works well in my experiments), which led to even better results.

I later found out about DLTV which made the agent discover new behaviors, but performed worse than previous experiments overall. Inspired by it, I tried something I did not find in previous works and got the best results out of all my previous experiments :

At each time step, compute an exploration_score as the ratio of "intra-action variance" over "inter-action variance" (rendered latex equation). I then take N/exploration_score samples from each distribution, and pick an action as described above. (more details at the end of this post)

For anyone reading this, I have a few questions :

1. Are you aware of any previous work I missed that tries similar exploration policies with distributional RL (interpolating between Thompson sampling and the greedy policy)
2. Most papers I found about learning from multiple exploration policies seem to be in the context of multi-actor parallelization. Is there any novelty in randomizing the policy parameters at the start of each episode, especially in the single-actor case ?
3. Is any part of what I'm doing worth the time it would take to quantitatively evaluate it ? I've been doing it mainly for learning and fun and have only qualitatively evaluated it so far. However, if there's a chance I can contribute to the field, I'll gladly make some time to compare it to published papers on ALE.

---

A few more details

I actually track a moving average and standard deviation of the exploration score, which lets me shift/rescale its values to a target average and standard deviation, and divide N by the shifted/rescaled value. I initially started with a target average of 1 and standard deviation of 1 as well (which gave good results), then tried randomizing these parameters at the start of each episode as well. This led to a lot more diversity in the policies and even better results.

Since this worked so well, I additionally randomized the noise strength in the NoisyNet layers.

Overall, this made the agent a lot more robust to deviating from what it considers to be the optimal trajectory, and allowed it to learn complex behaviors previous iterations were never able to learn (e.g. taking a few steps back to gain momentum, waiting for good cycles, or dodging hammer bros)

---

Watch it learn

For anyone interested, I made a live stream of the training in progress with graphs and some more details on the experiments I'm running. The current training run was started ~2.5 days ago. The agent has finished and unlocked levels up to 5-1, and is currently learning 5-2.

---

A lot more details

::: spoiler Long text hidden, click to expand Available actions : The agent does not have access to the up and down buttons, the available actions only use left, right, A and B.

Adding the down button would double the total number of actions (because down can be pressed on top of all available actions).

Reward function : It mainly consists of reward(t) = max(0, x(t) - previous_best_x) + a larger reward for beating a stage. I had to tweak the scaling of both components.

I initially had penalties for time and death, but one made the agent suicidal in front of hard-to-overcome obstacles, while the other made it fear them too much and hug the left side of the screen. Removing both proved to increase the performance.

One trick that seems to help with most '*-3' levels (which have a lot of void to fall into) was to hold the reward while the vertical velocity of Mario is negative (meaning it is falling). Without this trick, the agent would sometimes get stuck learning to jump the farthest it can into the void.

Stage scheduling : Each episode is one attempt on one level. At the start of each episode, a stage is randomly picked with probability proportional to 1/(number of times the stage was beaten) among the unlocked stages. Each stage is unlocked after the previous one has been beaten 30 times, with only 1-1 unlocked at the start of the training.

Available stages : The first iterations of the agent were unable to learn maze castles (4-3, 7-3 and 8-4), so I removed them all. The reward function will give rewards for the first path the agent tries, then the agent will be teleported back by the game and no reward is received until it finds the right path and gets past the point where the game teleported it back. I plan to test newer (better) versions of the agent on these stages only and see if mazes can be re-added to the pool.

I've also removed underwater stages (2-2 and 7-2). The agent can learn them fine, but the game dynamics are really different from all other stages and they're really boring to watch. Since I already removed a bunch of stages, I figured I could remove these as well but I may re-add them with mazes because beating every level is cooler than beating a cherry-picked selection.

Since 8-4 is the only stage that requires going down a pipe, I considered it was not worth it to add the down action and will likely never re-add it to the pool, which would unfortunately be really anti-climactic...

Replay buffer warm-up : After initially using the standard approach of filling the buffer with transitions sampled from a random policy before training the neural net, I came-up with a "soft warm-up" scheme in which the first gradient updates happen after only 2000 transitions, but initially happen every few thousand transitions and gradually become more frequent until the replay buffer is full. Together with my custom exploration policy, this works very well : the agent very quickly starts behaving similar to a "right + random button" policy before learning to actually jump and run.

Custom n-step bootstrapping : When I initially implemented n-step bootstrap targets, I initially used n=3 from the Rainbow paper, noting the same instabilities as the paper did for higher n values. I then found the Retrace(\lambda) paper which seems to successfully address this by increasing n until the online network disagrees with the action choice from a stored transition. This makes n larger where the replay buffer data is on-policy, and smaller when it becomes off-policy. Since my GPU is already maxed and the training is already slow (20.8t/s when real-time is 20t/s) I could not afford the additional computations (building a training sample (s(t), a(t), sum(r(t+0..n)), s(t+n)) needs up to n_max transitions to go through the online network).

I'm trying to achieve similar sample efficiency gains by using cheaper alternatives as proxies for "how off-policy is a given transition" : I'm using the number of times a transition has been sampled, with n = int(max(n_min, n_max * k**times_sampled)) ; 0<k<1. The currently running experiment uses n_max=14, n_min=1 and k=1/1.3. I'm pretty sure it helps early in the training, and it does not collapse like a constant n=14 does

Stream setup : As I said, this is something I do for my own fun, and I really wanted to be able to see the agent learn in real time. The code runs a separate process, to which frames from training episodes are sent in a queue. The process then sends the frames as raw RGB24 to an local UDP socket, to which GStreamer connects and encodes the stream. With a simple MediaMTX configuration, I can manage the Gstreamer process and have the stream available through WebRTC on my LAN.

Then I figured someone else might have fun watching this, so I added a line to my MediaMTX config to send the stream to twitch and youtube. The overlay is a headless browser displaying custom HTML/JS (using d3.js for the graphs) piping raw frames to ffmpeg. GStreamer handles compositing the two streams together into the side-by-side view.

:::

We are Student One Causal Networks. We have built an exhaustive statistical enumeration engine that runs full parameter-space searches across technical indicators, timeframes, and asset classes. The output is not a signal or a strategy recommendation — it is a structured dataset of statistically validated configurations. A neutral compute layer, not an opinion. The current build is functional for human users. What we are doing now is making it natively consumable by AI agents — proper tool definitions, OpenAPI spec, machine-readable outputs, the full stack so an agent can invoke our engine mid-conversation and return a real computed answer instead of a guess. We are in advanced discussions with one of the top three AI companies in the world to integrate this infrastructure into their agent ecosystem. We cannot name them publicly at this stage, but we will share the full picture privately with anyone serious enough to reach out. We are looking for engineers who have worked with agent-tool protocols — MCP, function calling, GPT Actions, tool-use APIs, or anything that lets a model call external compute during inference. This is a founding-stage equity role. No salary. The kind of opportunity that either makes sense to you immediately or does not. Interns with relevant skills are also welcome to apply. studentone.tech Drop a comment or message directly. Tell us what you have built.

The AI boom is more likely a marker of the end of the 50-year digital boom than the start of a new wave of innovation.

I know this is probably a long shot, but I'm not sure where else to ask so I'm going to take a shot.

I've designed and abstract board game (think chess, shogi, go, etc) and have completed coding the rules for play against an AI player, however getting the actual AI to be good is a whole other problem.

I would love if someone who is experienced in ML would be interested in collaborating on this open source project.

The game is strictly a hobby project, with absolutely no plans for monitization or anything. Currently it's playable in the browser against AI (no multiplayer yet set up) at: https://greenants.github.io/Amalgam_Webgame/

GitHub Repo: https://github.com/GreenAnts/Amalgam_Webgame

Disclaimer: I've mostly used AI to code this project, as I'm a pretty novice programmer. Obviously that's controversial, so I want to make that clear - but remember this is simply a hobby project, and is a way for me to get my board game design digitized and actually played by others. The code will likely be a bit on the messy side, but I think for the most part the ML coder would only be interacting with the controller - so shouldn't be too much of a factor.

From my limited understanding, the actual search depth and complexity of the game is quite high, far higher than chess, so it's been quite hard for me to try and get this set up even with the help of AI coding with hueristics.

If you are interested in in the project at all, I'm always looking for help to farther this project - as I've been working on the board game itself (on and off) for more than 10 years.

The GitHub Repo listed above (in the README.md) has a graphical rulebook as well as a video tutorial linked for you to learn the rules and get an idea of the game complexity if you are interested.

Like I said, I know this is a long shot, and unlikely anyone will be interested, but I figured I'd give it a shot :)

cross-posted from: https://lemmy.ca/post/58902343

https://theonion.com/tips-for-using-ai/

Deboo — JWG Dialogue Mode, engage! Deboo: Explain machine learning 🤓🤓 JWG: Imagine giving a computer a giant basket of examples and whispering, “Figure out the pattern hiding in here.” The machine squints (metaphorically), pokes around the data, adjusts a zillion tiny dials ins...

cross-posted from: https://lemmy.sdf.org/post/42723239

Archived

Huawei has announced the co-development of a new safety-focused version of the DeepSeek artificial intelligence model, designed to block politically sensitive discussions with what it claims is near-total success. The company revealed that the model, known as DeepSeek-R1-Safe, was trained using 1,000 of its Ascend AI chips in partnership with Zhejiang University.

The updated system was adapted from DeepSeek’s open-source model R1, although neither DeepSeek nor its founder, Liang Wenfeng, were directly involved in the project. Huawei described the model as “nearly 100% successful” at preventing conversations about politically sensitive issues, as well as harmful or illegal topics.

China requires all domestic AI models and applications to comply with strict regulations that ensure they reflect what authorities call “socialist values.” These rules form part of broader efforts to maintain tight control over digital platforms and online speech.

[...]

Highlights include: GIF decode and encode for imgcodecs, improved PNG and Animated PNG files handing, animated WebP Support, and especially the new HAL for RISC-V RVV 1.0 platforms.